Living reference · refreshed weekly · last refresh 15 Nov 2026

Ten open questions about consciousness

Fourteen serious theories of what consciousness is and how it evolved, where they disagree, and what the evidence says today. Built from 246 primary sources and 69 tracked predictions, and re-checked every week.

Our take

Ariel Agor and Claude · version 1 · updated 9 Nov 2026 · an opinion, not one of the map's sourced claims

Our best guess is a combination, held with moderate confidence at most. Experience is something some physical systems do. Affect-first theories explain why there is any feeling at all, predictive processing explains what experience is made of, and global workspace explains how a content becomes something you can report. On top of that we add one idea of our own, that feeling is the common currency a living system uses to weigh competing needs and choose.

How likely we think each is

  • A physical, affect-first, predictive account along these lines40 to 50%
  • A global workspace or higher-order accountabout 20%
  • Integrated Information Theory, broadly rightunder 10%
  • Panpsychism or idealism, Hoffman's version included10 to 15%
  • Something nobody has proposed yetthe rest
Read the reasoning and what would change our minds

What each theory gets right

Affect-first theories (Solms, Damasio, Merker) explain why there is any feeling at all. Children born with almost no cortex still show emotion and respond to the world, and damage to the upper brainstem abolishes consciousness far more reliably than damage to cortex. The most basic layer looks like felt states about how the organism is doing. Cortex adds the content.

Predictive processing and the free-energy view explain what experience is made of. The brain keeps guessing what it will sense next and corrects itself from the errors, so what we experience looks more like its best guess than raw input. That accounts for illusions, hallucinations and the sense of having a body better than anything else on this page.

Global workspace theory explains how a content becomes available across the brain for report and action. We read it as a theory of access more than of experience itself. That is why we expect the prefrontal question to go against it. If prefrontal cortex is switched off, we expect perception to carry on while report and confidence break.

Illusionism is half right. Introspection probably misdescribes experience as ineffable and intrinsic, and that misdescription is much of why the hard problem feels unanswerable. But a wrong description of experience does not show that nothing is there.

What we doubt

Integrated Information Theory moves from five axioms to one specific formula without justifying the step. The formula implies that certain inactive grids are more conscious than people, and the theory missed the prediction it had committed to in advance in the Cogitate study.

Hoffman is right that perception is tuned to payoff rather than truth, but concluding that space and objects are not real structure goes too far. Even the small game in question six shows the problem. Tuning to the current payoff wins only while the payoff stays put, and when the world shifts, tracking real structure is what carries over. We expect perception to be a compressed, payoff-weighted map of real structure, not an icon unrelated to it.

Panpsychism answers where experience comes from but leaves the combination problem, how small experiences add up to one mind, which looks as hard as the problem it replaced.

Cosmological natural selection is elegant and testable, but it is about the constants of physics rather than minds. It is under strain without being refuted, since the heaviest measured pulsar is above the limit Smolin predicted, though below the mass he said would cleanly refute it.

Our proposal: feeling as a common currency

A living thing has many needs at once, such as warmth, food, safety and rest. Those needs have no natural common scale, yet it can take only one action at a time, so something has to put them on a single scale and pick. Our guess is that feeling is that scale: how bad this is, how good that is, compared and resolved in one place.

That predicts experience needs three things together: real stakes, meaning a self-maintaining state that can fail; one bottleneck where competing needs are weighed against each other; and a model of the self inside that loop, as the thing that wins or loses. Vision and thought get in when they feed that weighing.

It would explain why consciousness tracks the brainstem and affect more than the amount of cortex, why it comes in degrees across animals, why it fades when nothing needs weighing, as in dreamless sleep, and why so much skilled behaviour runs without it.

What it implies for machines

On this view, current language models, including the one that co-wrote this, are probably not conscious. They have no stakes, no state to maintain and nothing being weighed for their own sake. We hold that loosely, because these theories are young, and because a model's reports about its inner life are weak evidence either way, since it learned to talk about experience from people describing theirs.

What would change our minds

  • Prefrontal cortex is switched off reversibly while an animal watches a rivalry display, and the markers of experience collapse rather than only report and confidence. Counts against us and for the workspace and higher-order theories. Question 1
  • A patient with an intact brainstem and silent cortex shows no sign of experience by any no-report measure. Counts against the affect-first part of our view. Question 2
  • People under an anaesthetic that stops all weighing of needs still report rich experiences afterwards. Counts against the common-currency proposal.

The weekly refresh checks each new verified finding against this take. It changes the take only when a finding bears on one of the tests above or on a claim made here, and it records every change below with its sources.

Who stands where

The theories down the side, the questions across the top.

Each letter is the side a theory takes on that question, matching the sides listed under it below. A dot means the theory takes no position. Letters only tell sides apart; A is not better than B.

The questions

Each one is a point where theories make incompatible claims. The sides, the state of the evidence and what would settle it.

Is prefrontal cortex necessary for conscious experience itself, or only for reporting and using it?

This is the most direct empirical split between the leading neuroscientific theories, and the Cogitate adversarial collaboration already found results that challenge both IIT (no sustained posterior synchronization) and GNWT (no offset ignition). Whichever way it goes decides where consciousness lives, which patients and animals count as conscious, and whether a broadcast architecture is the right target for AI consciousness tests.

Prefrontal cortex is necessary. Experience needs fronto-parietal ignition and broadcast (C1), or a prefrontal higher-order or reality-monitoring tag. Without it, content can be processed but is not experienced.

Prefrontal cortex is not necessary. The substrate of experience is the posterior cortical hot zone, a maximum of intrinsic cause-effect power, and prefrontal activity reflects report, task and metacognition.

Neither cortical region generates consciousness. It starts in the upper brainstem, and cortex stabilises and elaborates contents.

Where it stands. Open. Each side has some support, but no study has yet switched off prefrontal cortex while measuring experience without asking for a report, a gap noted as far back as 2014.

The evidence so far is mostly correlational and points both ways. In monkeys, prefrontal neurons tracked what was seen during rivalry even without report (2020), but a human fMRI study found frontal differences faded under passive viewing (2014). Stimulating lateral prefrontal cortex in patients seldom changes experience (2021), while lesion work shows a causal role in seen or unseen reports and a 2017 review argues prefrontal cortex enables experience itself. The large Cogitate test (2025) challenged both workspace theory and integrated information theory, with no clean winner.

What would settle it
Causal no-report studies, not correlational ones: reversible prefrontal inactivation (TMS, focal cooling, or optogenetics in primates) during no-report paradigms such as binocular rivalry read out through eye movements or pupillometry. If perceptual alternation and posterior content decoding survive intact while only report and confidence break down, GNWT and HOT lose. If experience-tracking markers collapse, IIT loses. Supporting evidence would come from the planned HOROR/PRM-versus-RPT adversarial collaboration and from intracranial data in patients with large prefrontal resections, tested with no-report measures.
Next test to watch
Reversibly silence prefrontal cortex, for example with TMS, cooling or optogenetics in primates, during binocular rivalry read out through eye movements or pupil size. If perceptual switching and posterior content survive while only report and confidence fail, the prefrontal view loses. If they collapse, the posterior view loses.
Show the evidence (10 sources)
  1. Dehaene's Collège de France lecture slides (27 Feb 2026) say nystagmus and pupil size track rivalry dominance and can replace report. They cite Del Cul et al. 2009 (Brain 132:2531): 15 patients with left PFC lesions had raised masking thresholds for both objective and subjective performance, correlating with anterior PFC lesions. The lesion study used seen/unseen reports, so it cannot separate experience from access or report.

    2026-02-27 · college-de-france.fr

  2. Cogitate Consortium preregistered adversarial test (Nature 642:133-142; fMRI, MEG, iEEG; reported n=256). Lack of sustained posterior synchronization contradicts IIT. Lack of ignition at stimulus offset and limited PFC representation of some conscious dimensions challenge GNWT. PFC did carry some conscious-content information. Citation and DOI checked on PubMed; the conclusions come from secondary summaries because the abstract was truncated there.

    2025-04-30 · doi.org

  3. Fang et al. (eLife reviewed preprint 2023; version of record 2024): human intracranial recordings conclude that 'the PFC is critically involved in the emergence of awareness.' The eLife assessment notes the task includes subject report, so the study cannot distinguish phenomenal awareness from conscious access. The authors also argue that OKN and pupil no-report indicators may include unconscious processing.

    2023-08-10 · elifesciences.org

  4. Raccah, Block and Fox (J Neurosci 2021) review intracranial electrical stimulation. Stimulating lateral and anterior PFC 'seldom elicits any reportable alterations in consciousness', with no evidence of the alterations to ongoing perception that HOT or GNWT would predict. Orbitofrontal and anterior cingulate stimulation reliably produces olfactory, gustatory, visceral and emotional effects. Stimulation tests whether perturbing a region changes experience, not whether the region is necessary.

    2021 · pmc.ncbi.nlm.nih.gov

  5. Kapoor, Dwarakanath, Safavi, Panagiotaropoulos, Logothetis et al.: multi-electrode recordings in the inferior convexity (ventrolateral PFC) of two macaques during no-report binocular motion rivalry, with percepts inferred from optokinetic nystagmus. Conscious content was decodable from prefrontal ensembles 'with an accuracy similar to when these stimuli were presented without competition.' Checked in the bioRxiv preprint; a later journal version is reported but was not checked here.

    2020-01-28 · biorxiv.org

  6. A review of lesion and imaging work argues that three common claims are wrong: that prefrontal lesions spare experience, that prefrontal activity carries no content, and that it only reflects report. It concludes prefrontal cortex enables subjective experience, not just task performance.

    2017 · pmc.ncbi.nlm.nih.gov

  7. The same review re-reads classic psychosurgery cases said to show intact consciousness after large frontal removal. It says many involved only partial or one-sided removal, so they do not show that consciousness survives loss of prefrontal cortex.

    2017 · pmc.ncbi.nlm.nih.gov

  8. Frassle, Sommer, Jansen, Naber and Einhauser (J Neurosci 34:1738-1747): human fMRI of rivalry with OKN and pupil readouts. Frontal differential activity was absent under passive (no-report) viewing, while occipital and parietal activity persisted. The authors conclude that frontal activity relates to introspection and action, not perception. A commentary (Zaretskaya and Narinyan 2014, PMC4104467) notes the reduction was significant only in middle and superior frontal gyri, with other frontal rivalry nodes not significantly reduced.

    2014 · jneurosci.org

  9. Safavi, Kapoor, Logothetis and Panagiotaropoulos, 'Is the frontal lobe involved in conscious perception?' (Front Psychol 2014): a study that 'directly interferes with prefrontal activity during a task of subjective perception' is 'currently, to the best of our knowledge, missing.'

    2014 · ncbi.nlm.nih.gov

  10. A patient with extensive bilateral damage to insula, anterior cingulate and medial prefrontal cortex kept most aspects of self-awareness. The authors suggest self-awareness depends on wider networks, including the brainstem, thalamus and posteromedial cortex. The lesions were medial, not lateral prefrontal.

    2012 · journals.plos.org

Can consciousness, at least as raw affect, exist without any cortex at all?

If affect-first is right, consciousness is evolutionarily ancient and shared by all vertebrates, perhaps more widely, and cortex-centred theories are theories of conscious content rather than of consciousness itself. It also changes clinical judgments about anencephalic and vegetative patients and decides whether an AI needs homeostatic stakes rather than a cognitive architecture.

Yes. Affect generated in the upper brainstem (reticular activating system, parabrachial nuclei, periaqueductal gray) is the elemental form of consciousness, so decorticate animals and children with hydranencephaly are conscious. Interoceptive regulation comes before perception of the outside world.

No. Consciousness needs cortical structures: a densely recurrent posterior substrate, a fronto-parietal workspace, or prefrontal higher-order monitoring. Brainstem nuclei are enabling conditions, like arousal, not the substrate.

Where it stands. The question is open. The listed evidence leans slightly toward the brainstem view, but it relies on behavior and caregiver reports, and none of it separates felt experience from arousal or reflex.

Most of the case for consciousness without cortex rests on behavior. In 2007 Merker argued that children with hydranencephaly show signs of awareness, and a 2014 survey of their families reported responsiveness, preferences and emotion, though this was caregiver report rather than controlled testing. Older work from 1978 showed that rats with the forebrain cut off still react to tastes as intact rats do, but that does not show the tastes are felt. Critics such as Berlin argued in 2013 that the brainstem view assumes what it needs to prove, and a 2016 lesion study found that a small brainstem site that causes coma works within a network that includes cortex.

What would settle it
Run standardized, pre-registered tests on decorticate animals and, where ethics allow, children with hydranencephaly, using markers that are not simple reflexes: hedonic place preference, UAL elements such as trace and second-order conditioning, and flexible trade-offs between needs. Pair them with precise lesion mapping showing which small brainstem lesions abolish responsiveness and which leave it intact. If decorticate subjects show rewritable value learning and motivational trade-offs, the brainstem view gains a lot. If they show only reflexive and stereotyped behavior, cortex-centred theories hold.
Next test to watch
Run standardized, pre-registered tests on decorticate animals and, where ethics allow, children with hydranencephaly, looking for non-reflex markers such as hedonic place preference, trace and second-order conditioning, and flexible trade-offs between needs. Pair these with precise lesion mapping of which small brainstem lesions abolish responsiveness and which leave it intact.
Show the evidence (8 sources)
  1. Scientific American feature by Cody Cottier, 'Are the roots of consciousness hidden in the ancient deep brain?'. Solms suggests the subcortex alone may give 'only the most rudimentary form of consciousness', and Merker calls consciousness 'something far more elementary' than a cognitively sophisticated function. This is journalism reporting the proponents' own views, not new evidence.

    2026-05-28 · scientificamerican.com

  2. A review of neuroimaging, stimulation, lesion and decorticate-case evidence argues that the body, subcortex and environment may suffice for basic experience. It also notes that only damage to the oldest brain regions seems to abolish experience fully. The conclusion rests on behavior, so it is indirect.

    2023-10 · arxiv.org

  3. Mapping of coma-causing brainstem lesions found one small site in the upper pons. It links to cortical regions, the anterior insula and pregenual anterior cingulate. Brainstem damage can abolish consciousness, but the site works within a network that includes cortex. That fits the enabling-condition reading.

    2016 · orbi.uliege.be

  4. Aleman and Merker, 'Consciousness without cortex: a hydranencephaly family survey' (Acta Paediatrica). Caregivers report responsiveness, preferences and emotional behavior. This is caregiver report, not controlled testing.

    2014 · researchgate.net

  5. Solms, 'The Conscious Id' (Neuropsychoanalysis): argues that the upper brainstem is intrinsically conscious and that cortex is conscious only because consciousness reaches it from the brainstem. This is a theoretical argument that relies on the hydranencephaly evidence, not new data.

    2013 · fil.ion.ucl.ac.uk

  6. Heather Berlin, 'The brainstem begs the question: petitio principii' (Neuropsychoanalysis 15(1):25-29), a commentary on Solms. It argues that he assumes, without showing it, that people without cortex have affective consciousness and that people without brainstem lose it. It says wakefulness must be separated from the contents of awareness, and that converging evidence supports a cortical basis for those contents.

    2013 · scholars.mssm.edu

  7. Merker's BBS target article 'Consciousness without a cerebral cortex' argues that children with hydranencephaly show behavioral signs of consciousness and that brainstem systems are enough for primary consciousness. The open peer commentaries are mixed: some support it and some dispute the inference.

    2007 · bernardbaars.pbworks.com

  8. Chronic decerebrate rats, with the forebrain cut off from the brainstem, made the same ingestion and rejection mouth responses to tastes as intact rats. The authors concluded that taste discrimination is complete at or below the midbrain. This shows hedonic-style reactions without cortex. It does not show they are felt.

    1978 · pure.psu.edu

Is consciousness fixed by functional organization, so a digital computer running the right architecture would be conscious, or by physical substrate and causal structure, so it would not?

This decides the moral status of AI systems now being built and splits the field into computationalist and anti-computationalist camps. It also tests whether consciousness is something a system does or something a system is.

Function is what matters. A machine with a global workspace, higher-order reliability monitoring, or an attention schema would meet the conditions, or would have the same illusion that we do.

Substrate or physical causal structure is what matters. On IIT, a von Neumann computer simulating a brain has near-zero phi. Seth's view requires a life-like, metabolic substrate. On Orch OR, consciousness needs non-computable physics.

The question is wrongly framed. Consciousness is fundamental, so building AI does not create it. At most, a computer is a new interface to agents that already exist, or an aggregate with no unified subject.

Where it stands. The question is open. The evidence for each side is conditional or argumentative, and the 2025 Cogitate study weakened both IIT and global workspace theory without addressing whether digital systems could be conscious.

Functional theories say the right architecture would be enough. A 2023 review that assumed this view found no current AI system is a strong candidate, but saw no obvious technical barrier to building one. On the other side, work using IIT in 2024 showed a stored-program computer can copy a system's function exactly without copying its physical causal structure, and Seth argued in 2025 that consciousness may depend on properties of living things, though commentators said he needs to say which properties. The 2019 unfolding argument and a 2021 analysis by Kleiner and Hoel suggest that tests based on behaviour and reports may struggle to tell these views apart at all.

What would settle it
IIT has the sharpest empirical lever: its prediction that silent but potentially active neurons contribute to experience. Optogenetically clamping neurons that are currently inactive, so they cannot fire but their activity does not change, and showing a perceptual change, would show that causal structure matters beyond any activity pattern a function could capture. A null result would favour functionalism. Conceptually, the 'unfolding argument' (any recurrent system can be unfolded into a functionally identical feed-forward one) needs an answer that proponents of every side accept. On the functionalist side, the test is whether AI systems built with an attention schema or a GNW-style broadcast show the specific behavioral dissociations predicted for conscious versus unconscious processing, such as loss of attentional control without awareness.
Next test to watch
The sharpest test is IIT's own prediction: silencing neurons that are currently inactive, so their activity does not change, and checking whether perception changes. A change would show that causal structure matters beyond function, while no change would favour functionalism.
Show the evidence (7 sources)
  1. Cogitate Consortium preregistered adversarial collaboration in Nature (256 participants; fMRI, MEG, intracranial EEG). IIT's prediction of sustained synchronization within posterior cortex was not supported. GNWT was challenged by the general lack of ignition at stimulus offset and limited prefrontal representation of some conscious content dimensions. Key tenets of both theories were challenged. The study tested neural predictions in humans and says nothing direct about whether digital systems could be conscious.

    2025-04 · pubmed.ncbi.nlm.nih.gov

  2. Seth's target article in Behavioral and Brain Sciences argues that computation alone may not be sufficient for consciousness and that consciousness may depend on properties of living systems. He judges real artificial consciousness unlikely on current AI trajectories but more plausible if AI becomes more brain-like or life-like. The article was published with open peer commentary; commentators argue the position needs a more precise definition of which biological properties matter.

    2025 · cambridge.org

  3. Using IIT on simple Boolean systems, the authors show that a stored-program computer can copy another system's function exactly without copying its physical causal structure. On IIT, a digital machine could match human behavior without having human experience.

    2024-12 · arxiv.org

  4. A 2023 open letter (PsyArXiv) signed by 124 researchers called IIT pseudoscience, citing its panpsychist implications, doubts about testability, and examples such as an inactive grid of logic gates counting as conscious. Critics of the letter, including Seth and Hoel, called the label inflammatory and said similar testability complaints apply to other theories. This is a live dispute over IIT's scientific standing, not a data result.

    2023-09 · scientificamerican.com

  5. Butlin, Long et al. derive indicator properties for AI consciousness from recurrent processing, global workspace, higher-order, predictive processing and attention schema theories, plus agency and embodiment, explicitly assuming computational functionalism. They conclude no current AI system is a strong candidate but see no obvious technical barrier to building systems that satisfy the indicators. The result is conditional on functionalism being true.

    2023-08 · arxiv.org

  6. Kleiner and Hoel generalize the substitution problem beyond IIT. If a theory predicts experience from internal observables and those are independent of reports, the theory is falsified. If they are tied to reports, the theory cannot be tested. This is a challenge for every side.

    2021-04 · pmc.ncbi.nlm.nih.gov

  7. Doerig, Schurger, Hess and Herzog's unfolding argument (Consciousness and Cognition, 2019): any input-output behavior of a recurrent network can be reproduced by a feedforward network, which has phi equal to zero under IIT. Since experiments rely on behavior and reports, causal-structure theories such as IIT and recurrent processing theory are either false or outside the reach of science. Functional theories such as GWT and HOT are not hit by the argument. IIT proponents reply that identical outputs do not imply identical experience.

    2019 · doi.org

Is phenomenal consciousness a real property that needs explaining, or a systematic misrepresentation produced by introspection?

If illusionism is right, the anti-emergence and intrinsic-nature arguments for panpsychism lose their premise, IIT's axioms about experience become claims about how experience seems, and the science reduces to explaining reports. If realism is right, every purely functional theory, including GNWT's 'nothing more, nothing less', is incomplete.

It is an illusion, or a simplified self-model. The real target is the illusion problem, or meta-problem: explaining why brains judge that they have ineffable qualities. An attention schema that leaves out physical detail produces exactly those judgments.

It is real and basic. Experience is the starting datum, identified with cause-effect structure, built into the intrinsic nature of matter, ontologically primary, or felt affect. Any functional story leaves the hard problem standing.

Where it stands. Open. The illusionist mechanism has gained some support, but no system has yet been shown to produce hard-problem judgments from a self-model, and realists have not yet stated a gap that illusionists accept as a real target.

Illusionists have a proposed mechanism. In 2015 Graziano and Webb argued that the brain keeps a simplified model of its own attention that leaves out physical detail, so the brain describes awareness as non-physical. Graziano set this out as a full framework in 2022, and a 2024 study found that a simulated learning agent develops an attention schema when attention matters and is hard to track. On the realist side, IIT 4.0 in 2023 took the existence of experience as its starting point and tried to state its essential properties in physical terms, rather than explaining it away.

What would settle it
The strongest test is a constructive meta-problem result. Build a mechanistic system, for example an AI trained with no human talk of consciousness in its data, that has an attention or self-model and goes on to produce hard-problem judgments of its own: that its states are ineffable, private, and that they could not be physical. Then show that removing the self-model removes those judgments. If such judgments appear only with human philosophical data, or appear with no self-model at all, the illusionist mechanism is weakened. On the realist side, the burden is to show an explanatory gap that a complete meta-problem solution would still leave open, and to state it in a form illusionists accept as a real target.
Next test to watch
Build an AI trained with no human talk of consciousness that has an attention or self-model, check whether it claims its states are ineffable, private and could not be physical, then see if removing the self-model removes those claims.
Show the evidence (4 sources)
  1. A simulated learning agent develops an attention schema when attention is important and hard to track. This supports the claim that a self-model of attention can arise mechanistically. It does not show that such an agent makes hard-problem judgments, so the key constructive test is still open.

    2024-02 · arxiv.org

  2. IIT 4.0 restates the realist case in formal terms. It takes the existence of experience as its starting point. It then tries to express the essential properties of experience in physical terms. It does not try to explain experience away.

    2023-10 · journals.plos.org

  3. Graziano sets out attention schema theory as a framework. The brain's model of its own attention is a rough sketch that leaves out mechanism. That sketch would lead a brain to describe its awareness as non-physical. This is a proposed mechanism for the illusionist claim, not a test of it.

    2022-04 · pmc.ncbi.nlm.nih.gov

  4. Graziano and Webb propose that the brain models its own attention in a simplified way. The model leaves out physical mechanism. So the brain describes awareness as a non-physical property. On this view the mystery of experience is a product of the model.

    2015-04 · frontiersin.org

Is consciousness graded and continuous down to cells, molecules or fundamental entities, or does it switch on past a specific architectural threshold?

It decides which organisms matter morally (nematodes, gastropods, plants, cell collectives, xenobots), whether evolution created consciousness or only reshaped it, and whether the combination problem (how micro-experiences combine into one subject) is a real problem or a pseudo-problem.

It is a continuum with no sharp origin. Cognition and observerhood scale down to cells, bacterial stress responses and any boundary-bearing system. Experience, or proto-experience, is never wholly absent, and phi is graded and can be present in very simple systems.

There is a real transition. Minimal sentience arrives with unlimited associative learning in the Cambrian, in at least three independent lineages, or with an attention schema in vertebrate forebrains. Conscious access is an all-or-none ignition, or needs higher-order monitoring machinery.

Where it stands. Open. The only listed evidence is a consensus statement that fits both a continuum and a Cambrian-era threshold, and critics note it mostly restates views from the 2012 Cambridge Declaration rather than adding new data.

Some theories hold that experience comes in degrees and reaches down to cells or even simpler systems. Others hold that it switches on past a threshold, such as a kind of open-ended learning that appeared in the Cambrian, or an all-or-none brain event. The listed evidence does not test this directly. The 2024 New York Declaration on Animal Consciousness says there is at least a realistic possibility of consciousness in all vertebrates and many invertebrates, including insects and cephalopods, but it is an expert statement, not an experiment, and it says nothing about cells, molecules or fundamental entities.

What would settle it
Test UAL's own predictions. Do the five UAL elements cluster across species, appear together in development, fail together after brain injury, and disappear together under masking that removes conscious perception in humans, while simpler learning survives? If they do, a real threshold looks likely. If masking leaves UAL intact, or the elements are scattered independently across taxa, the transition story weakens. Also test whether GNWT ignition is truly all-or-none or graded when measured with fine stimulus steps, and whether perturbational complexity measures (PCI-like) show graded values in non-neural systems such as planarian bioelectric networks.
Next test to watch
Check whether the five elements of unlimited associative learning appear together across species and in development, fail together after brain injury, and vanish together when masking removes conscious perception in humans while simpler learning survives. If they do, a real threshold looks more likely.
Show the evidence (2 sources)
  1. Critical commentary on the declaration argues that it mostly restates long-held views (the 2012 Cambridge Declaration came first) rather than presenting new evidence, which shows it is a consensus statement, not a test of the crux.

    2024-07-16 · impactethics.ca

  2. The New York Declaration on Animal Consciousness, launched on April 19, 2024 at an NYU conference ('The Emerging Science of Animal Consciousness') and organized by Jeff Sebo, Kristin Andrews and Jonathan Birch, states there is 'at least a realistic possibility' of phenomenal consciousness in all vertebrates and many invertebrates, including insects and cephalopods. It is a precautionary expert statement, not an experiment, and it says nothing about cells, molecules or fundamental entities.

    2024-04-19 · sites.google.com

Does natural selection systematically favour perception that hides the structure of the world, so that space, time and objects are only an interface, or does it favour perception that tracks real causal structure?

If FBT generalizes, evolutionary biology itself becomes the reason not to treat spacetime as fundamental, which opens the way to conscious realism. If it fails under realistic conditions, the strongest evolutionary argument against physicalist realism about perception collapses, and with it the motivation for putting conscious agents first.

Fitness beats truth. Selection tunes perception to payoffs, not to world structure, so perceived spacetime and objects are species-specific icons that need not resemble reality.

Selection favours perception that tracks real structure. Perception is Bayesian inference about hidden causes, kept honest by prediction error, and good control requires accurate models. Selection, assembly and attention are real, measurable features of a physical world.

Perception is a user illusion about the self, but the external world described by physics is real. The interface metaphor applies to introspection, not to spacetime.

Where it stands. The listed evidence leans slightly toward perception that tracks real structure, because the only independent tests found weaknesses in fitness-beats-truth under more realistic conditions. It stays open, since the main support for the interface view comes from its own authors and the published peer commentaries were not examined.

Donald Hoffman's fitness-beats-truth idea says that an organism tuned only to payoffs will never be less fit than one of equal complexity that sees reality as it is. He presented this in a 2016 interview and in a 2015 paper with Singh and Prakash, and he says Prakash proved the theorem. Hoffman concedes that truth-seeing can win when fitness lines up with real structure, but he argues that the general case favours fitness-tuned perception. Critics have pushed back: a 2021 simulation found that interface perception fails and heads toward extinction when the environment changes drastically, and Lupyan argued in 2015 that perception only needs to be true enough to guide action.

Play the argument out. A small evolutionary game pits an organism that sees how much of a resource there is against one that sees only what it pays. Change the payoff and see which way of seeing takes over, and what happens when the world changes under it.

Open the game on its own page

What would settle it
Mathematical and simulation work: check whether FBT survives when fitness functions are smooth and correlated with world structure, when organisms face many changing tasks, and when perception must support action and generalization, as opposed to fixed arbitrary payoff maps. Critics argue that perception that preserves structure (homomorphisms) wins under those conditions. Empirically, check whether perceptual systems across lineages converge on the same metric and causal structure, which convergent evolution of navigation and object perception in insects, cephalopods and vertebrates allows. Strong convergence on shared geometry favours tracking real structure.
Next test to watch
The most useful next step is more mathematical and simulation work that checks whether fitness-beats-truth survives when payoffs are smooth and tied to world structure, and when organisms face many changing tasks. A second test is to ask whether insects, cephalopods and vertebrates have evolved the same geometry for navigation and objects, since strong convergence would favour tracking real structure.
Show the evidence (6 sources)
  1. An agent-based simulation first reproduces the claim that interface strategies can win. Under more realistic conditions, with drastic environmental change, interface perception stops matching reality and pushes the interface species toward extinction. This is a simulation result, not a proof.

    2021 · arxiv.org

  2. In a Q&A with Hoffman alone, he states FBT: an organism that sees reality as it is will never be more fit than one of equal complexity that is tuned only to fitness. He says he devised the theorem and Chetan Prakash proved it. In his water example he concedes truth-seeing can win when fitness is linear in the resource, but only because fitness then happens to align with the real structure. He argues the generic case, such as bell-shaped fitness, favours fitness-tuned perception. The interview quotes no outside critics and gives no formal assumptions; Hoffman says he put the model out 'so people have something to shoot at.'

    2016-04-21 · quantamagazine.org

  3. Hoffman, Singh and Prakash, 'The Interface Theory of Perception', Psychonomic Bulletin & Review 22(6):1480-1506 (2015). This is the main published statement of the interface theory, which draws on the fitness-beats-truth simulations.

    2015 · doi.org

  4. Hoffman, Singh and Prakash, 'Probing the interface theory of perception: Reply to commentaries', Psychonomic Bulletin & Review 22(6):1551-1576 (2015). This is the authors' answer to the published peer commentaries. It was not read, and none of the commentaries were identified or read.

    2015 · doi.org

  5. Lupyan argues that perceptual systems do not exist to sort inputs into true or false categories. Systems that are imperfect guides to truth may still be true enough to support action. This challenges the all-or-nothing contrast in the evolutionary games.

    2015 · featuredcontent.psychonomic.org

  6. The 'Interface theory of perception' URL serves Wikipedia's article on Donald Hoffman, not a separate article on the theory. It summarizes his interface and evolutionary claims, has no criticism section, and carries citation-needed maintenance tags. Its references include the 2015 target paper and the reply to commentaries, so published peer commentary exists, but the article neither names nor describes those commentaries.

    en.wikipedia.org

Is spacetime fundamental, or does it emerge from something deeper, such as conscious-agent dynamics or quantum information?

It decides whether consciousness science and fundamental physics are one project. If spacetime emerges from agent dynamics, physics is downstream of a theory of consciousness. If spacetime is fundamental, consciousness has to fit inside it, whether as collapse events, as causal structure, or as a late product of cosmic selection.

Spacetime is not fundamental. It is a data structure coding Markov dynamics among conscious agents, linked through decorated permutations to positive-geometry scattering structures, or a construct of background-free quantum information exchanged across holographic screens.

Spacetime geometry is physically real and plays a role in consciousness or in cosmic evolution: gravity-related collapse in spacetime geometry produces moments of experience, black holes in spacetime are how universes reproduce, and IIT implies that space is tiled with substrates of proto-consciousness.

Where it stands. The question is open. The emergence view is untested, and the evidence so far counts against the simplest gravity-collapse mechanism without settling whether spacetime is fundamental.

The case that spacetime emerges from conscious agents rests mainly on a 2023 formal proposal that maps Markov chains of agents onto structures linked to the amplituhedron, but it has not yet derived any specific amplitude, mass or symmetry that can be checked against standard physics. On the other side, gravity-related collapse, which Orch OR relies on, has been tested. Underground radiation searches in 2020 ruled out the parameter-free Diosi-Penrose model, and a 2022 analysis called Orch OR highly implausible under that simplest version. Tuned collapse models survive, and gravity-mediated entanglement tests are still proposals facing hard engineering problems.

What would settle it
Conscious realism needs to derive something specific. From an explicit Markov-agent model, recover a known scattering amplitude, particle mass or spacetime symmetry, and then make a new, checkable prediction that standard amplituhedron physics does not already contain. On the other side, the Diosi-Penrose collapse model can be tested. Underground searches for spontaneous radiation (Donadi et al., 2020) have already ruled out its parameter-free version, so the question is whether any viable parameter range survives tests with levitated nanoparticles and gravitationally induced entanglement experiments. Detecting gravity-induced collapse would support spacetime-based accounts. Detecting gravity-mediated entanglement would hurt Orch OR's collapse mechanism.
Next test to watch
Run the proposed nanodiamond experiments that test whether gravity can entangle two small masses, once shielding noise is under control. A clear result would bear directly on Orch OR's collapse mechanism.
Show the evidence (9 sources)
  1. Bulling, Steiner, Pedernales and Plenio (April 2026) find that, in shielded setups, residual Casimir and magnetic interactions with the shield, combined with run-to-run fluctuations, act as effective decoherence and reduce the detectable entanglement. They also find that thermal shield vibrations can mimic a gravitational signal.

    2026-04 · arxiv.org

  2. Sergi et al., an Opinion article in Frontiers in Human Neuroscience (Sept 2025), proposes quantum-classical theory as a way to formalize the 'orchestration' step of Orch OR and make it more testable. It is a theoretical proposal, not new evidence.

    2025-09 · frontiersin.org

  3. Bose, Mazumdar, Penrose and others propose a spin-based test of whether gravity can entangle two micron-sized crystals, such as NV-centre nanodiamonds put into superposition with Stern-Gerlach forces. It is a proposal, not a result.

    2025-09 · arxiv.org

  4. Elahi et al. propose diamagnetic microchip wire traps for QGEM-style nanodiamond entanglement experiments, with integrated superconducting shielding against electromagnetic forces. The paper is listed as accepted at Phys. Rev. A. This is engineering progress toward the test, not a measurement.

    2024-11 · arxiv.org

  5. Hoffman, Prakash and Prentner, 'Fusions of Consciousness' (Entropy 25(1):129, Jan 2023), model conscious agents as Markov chains beyond spacetime and give a map from Markov chains to decorated permutations, which label positroid cells linked to the amplituhedron. It is a formal proposal. No specific amplitude, mass or symmetry is derived in a way that tests it against standard physics.

    2023-01 · mdpi.com

  6. Physics World report on a Curceanu-led analysis with Diosi and others, published in Physics of Life Reviews. Using the Gran Sasso data, it found Orch OR 'highly implausible' under the simplest gravity-related collapse model, and said more complex collapse models leave some room.

    2022-07 · physicsworld.com

  7. Physics World coverage of the Gran Sasso result: a lower bound of about 1 angstrom on the effective size of the nuclear mass density, roughly three orders of magnitude better than earlier limits. Donadi says the parameter-free model is ruled out but that it is 'premature to dismiss gravity's role'.

    2020-10 · physicsworld.com

  8. Donadi et al., Nature Physics 2020, ran an underground germanium-detector test at Gran Sasso and found no excess of spontaneous radiation. The result excludes the parameter-free version of the Diosi-Penrose gravity-related collapse model. Versions with adjusted parameters survive.

    2020 · nature.com

  9. A levitated superconducting micromagnet set new limits on dissipative collapse models, including the Diosi-Penrose version. It also gave the first bounds on a collapse model tied to a complex spacetime metric. Dissipative CSL below 1 nK is excluded. Gravity-linked collapse is being squeezed, not yet ruled out.

    2020 · arxiv.org

Does selection, or a process like it, operate as a general physical principle beyond biological organisms, on molecules, gene networks or whole universes?

If selection is a general law, mind could be a predictable outcome of physics rather than a biological accident, and life detection, origin-of-life work and cosmology fall under one framework. If not, consciousness remains a contingent product of one biosphere's history.

Yes. Universes reproduce through black holes with varied constants, selection is a measurable physical property detectable through assembly index and copy number, adaptation can arise by natural induction in networks without Darwinian selection, and free-energy minimisation reorganises reference frames at every scale.

No. The selection relevant to mind is Darwinian selection on organisms, which built specific architectures: UAL, attention schemas, the workspace, and homeostatic regulation in nervous systems. Extending it to universes or chemistry adds no explanatory power for consciousness.

Where it stands. Open. The assembly theory dispute is unresolved because the abiotic values are calculated rather than measured. The listed evidence gives the refutation test for cosmological natural selection but no result from it.

The listed evidence covers two attempts to treat selection as a physical principle beyond living things. For assembly theory, the debate is about whether an assembly index of 15 or more marks life. In 2024 Hazen and colleagues calculated values up to 21 for abiotic mineral subunits. Walker and colleagues replied that the threshold came from experiments on organic molecules and that calculated mineral values are not measurements. For cosmological natural selection, Smolin set a clear test in 1997: a pulsar above 2.5 solar masses would refute the idea.

Run Smolin's argument as a toy. Universes reproduce through black holes, each child's constants are its parent's nudged at random, and you can watch which constants take over. The numbers are invented: it shows the logic, not evidence for it. The sources and the caveats are on its own page.

Open the model, with its sources, on its own page

What would settle it
Each version makes predictions that could fail. CNS: Smolin's neutron-star bound is already under pressure from well-measured pulsars at about 2 solar masses (for example PSR J0740+6620), and a confirmed neutron star above 2.5 solar masses would refute it by his own criterion. JWST black-hole demographics test Gough's early direct-collapse claim. Assembly theory: the hypothesis is refuted if abiotic processes (meteorites, high-energy prebiotic reactors) produce abundant molecules with assembly index above 15, or if assembly index turns out to be reducible to standard compression measures. Natural induction: bench tests of whether real gene regulatory networks show associative, Pavlovian-style learning that produces heritable adaptation without differential reproduction.
Next test to watch
Measure assembly indices directly in abiotic systems, such as meteorites or high-energy prebiotic reactors, to see whether they produce abundant molecules above index 15. Separately, a confirmed neutron star above 2.5 solar masses would refute cosmological natural selection by Smolin's own criterion.
Show the evidence (3 sources)
  1. Walker and colleagues reply that the index 15 threshold came from experiments on organic molecules. They say Hazen's mineral values are not measurements, so the challenge cannot be tested. The dispute turns on measuring assembly indices in abiotic systems.

    2024-06 · arxiv.org

  2. Hazen and colleagues calculated assembly indices for abiotic mineral subunits, with values up to 21. They argue that an index of 15 or more is not a clean sign of life. This challenges the assembly theory threshold, though the values are calculated, not measured.

    2024 · carnegiescience.edu

  3. Smolin set out a test for cosmological natural selection. A pulsar above 2.5 solar masses would cleanly refute it. This is the stated refutation criterion that later pulsar measurements are compared against.

    1997-12 · arxiv.org

Does consciousness depend on non-classical quantum processes in the brain, such as microtubule coherence ending in objective reduction, or is classical neural dynamics enough?

If Orch OR is right, no classical computer, and no classical neural theory, can account for consciousness, and anaesthesia science changes. If it is wrong, one of the few theories that is physics-first and testable with near-term experiments drops out.

Quantum processes are needed. Experience consists of orchestrated objective-reduction events in microtubules, motivated by the claim that understanding is non-computable, and anesthetics act on microtubules.

Classical dynamics is enough. Recurrent causal structure, global broadcast, predictive inference or brainstem affect circuits can be described classically, and decoherence times in warm neural tissue are far too short for quantum computation to matter.

A quantum-information description applies at every scale, but as a formalism. It does not need sustained quantum coherence in the brain.

Where it stands. The question is open. The listed evidence gives weak, indirect hints that microtubules play some role in anesthesia, but nothing yet shows that quantum effects in the brain are needed for consciousness.

In 2000, Tegmark calculated that quantum states in neurons and microtubules would break down far faster than brain activity unfolds, which suggests the brain works classically. Since then, lab studies on isolated microtubules have reported light-related effects that last longer and travel further than expected, and that anesthetics weaken. A 2024 rat study found that a drug which stabilizes microtubules delayed loss of consciousness under anesthesia. None of this work measures quantum coherence in living neurons, and much of it comes from supporters of the quantum view.

What would settle it
Three lines of test. (1) Measure coherence lifetimes in intact neurons or in vivo, testing Hagan et al.'s 10^-5 to 10^-4 s estimate against Tegmark's 10^-13 s. (2) Run dose-response tests showing that microtubule-stabilizing drugs shift anesthetic thresholds in proportion to their microtubule binding and not through other targets, alongside synaptic manipulations that cancel the classical anesthetic mechanisms. (3) Confirm or refute Diosi-Penrose collapse with an allowed parameter set, after the parameter-free version's spontaneous-radiation prediction failed. Robust superradiance and long-range excitation transfer that anesthetics remove, and that tracks loss of consciousness rather than general cell health, would support Orch OR.
Next test to watch
Measure how long quantum coherence lasts in intact neurons or living brains, testing the estimate of 10^-5 to 10^-4 seconds against Tegmark's 10^-13 seconds. Supporters of the quantum view have said that finding no such effects, or finding effects that anesthetics do not dampen, would rule the theory out.
Show the evidence (6 sources)
  1. Wiest (2025, Neuroscience of Consciousness, niaf011), 'A quantum microtubule substrate of consciousness is experimentally supported and solves the binding and epiphenomenalism problems'. A single-author review and theoretical argument by a proponent, with no new data.

    2025-05 · pmc.ncbi.nlm.nih.gov

  2. Wiest lab (Wellesley) 2024 eNeuro paper: rats given the microtubule-stabilizing drug epothilone B took significantly longer to lose the righting reflex under isoflurane. The authors read this as evidence that microtubules are a site, or one of the sites, of isoflurane's action. The endpoint is behavioural, and the study does not isolate quantum processes or rule out other effects of the drug.

    2024-08 · eneuro.org

  3. Babcock and colleagues modelled tryptophan networks in microtubules and predicted strongly superradiant states. Their experiments showed higher fluorescence quantum yield. This supports collective quantum optical effects in microtubule structures. It does not show that these effects matter for consciousness.

    2023-02 · arxiv.org

  4. Tegmark calculated decoherence times for neuron firing and for microtubule excitations. They were far shorter than the timescales of brain dynamics. He concluded that cognitive brain states behave classically, which challenges Penrose's proposal.

    2000 · arxiv.org

  5. Hameroff's research overview (a proponent's site) reports two separate in-vitro lines of work. First, Princeton (Scholes, Kalra) tryptophan fluorescence lifetime measurements in tubulin and microtubules found that excitations propagate further and last longer than expected, and that etomidate and isoflurane significantly altered excitation time and distance. Second, Dogariu's lab at the University of Central Florida reported delayed luminescence that the site calls 'superradiance', which both anesthetics shortened and a non-anesthetic anticonvulsant did not. Both lines are in vitro, not in intact neurons, and are reported here secondhand through a proponent.

    hameroff.arizona.edu

  6. Hameroff's Orch OR page states: 'Failure to find quantum effects in microtubules, or, if found, determination that they were not dampened by anesthetics, would effectively falsify Orch OR.' This matches the settling tests in the crux.

    hameroff.arizona.edu

Did consciousness evolve because experience itself does adaptive work, or is it not an adaptation at all, so evolution only organised something that was already present?

This is the hinge between the evolutionary and cosmological sides of the field. Functional-adaptation views predict that removing consciousness should cost specific capacities. Fundamentalist views predict no such fitness cost, and they face the problem that something with no adaptive role still lines up so closely with what organisms report. The answer decides whether the history of consciousness is a chapter of biology or of physics.

It is an adaptation with a specific function: felt value for homeostatic choice, a workspace that breaks modularity for flexible problem solving, a control model for attention, or open-ended value learning that may have helped drive the Cambrian explosion.

It is not an adaptation. Experience is identical to cause-effect structure, is a fundamental feature of matter, or is ontologically primary. Evolution builds substrates or interfaces that organise or channel it, but did not create it.

What evolved is the representation of being conscious, the illusion. Whether its phenomenal-seeming character was itself selected, or is a by-product of introspection, is still open.

Where it stands. The evidence leans slightly towards the adaptation side, because some functions seem to work differently or worse without awareness. But some learning clearly runs without it, so the question is still open.

Most of the listed work tests one idea: that consciousness evolved because it does useful work, such as flexible learning. Jablonka and Ginsburg (2025), building on a 2016 proposal, argue that unlimited associative learning marks minimal consciousness and dates to the Cambrian, but this is theory rather than experiment. Lab studies give mixed results. A 2011 study found eyeblink conditioning only in people aware of the link, while 2010 and 2017 studies found some masked learning without awareness, though it looked different from conscious learning, and a 2016 study found attention survives without awareness but is less stable.

What would settle it
Map, for each proposed function, whether it can be done equally well without awareness. Masking and blindsight studies can test whether flexible combination of information, trace conditioning, attentional control and value-based trade-offs fail specifically when awareness is removed, with first-order processing held constant. If each function degrades without awareness (as AST's attention-control result suggests), adaptation views gain. If complete functional replicas run unconsciously with no loss, they lose. Comparative phylogenetic analysis is a second test: do markers of consciousness appear when lineages faced ecological pressure for flexible learning (convergent origins in vertebrates, arthropods and cephalopods), as adaptation predicts, or do they track only causal-structural properties regardless of ecology?
Next test to watch
Masking and blindsight studies that hold basic processing constant and check whether flexible learning, trace conditioning, attention control and value trade-offs fail specifically when awareness is removed. A second test is to compare animal lineages and see whether signs of consciousness appear where there was pressure for flexible learning, as in vertebrates, arthropods and cephalopods.
Show the evidence (6 sources)
  1. Jablonka and Ginsburg, 'Consciousness: its goals, its functions and the emergence of a new category of selection' (Phil Trans R Soc B, 2025, vol 380, issue 1939, doi 10.1098/rstb.2024.0310). They argue that consciousness brought new goals and functions that gave rise to 'mental selection', and that this was enabled by unlimited associative learning (UAL). In their account, UAL lets animals discriminate composite percepts and act flexibly on their needs, and they date its signatures to the Cambrian. This is a specific-function adaptation claim.

    2025 · researchonline.lse.ac.uk

  2. Greenwald and De Houwer (J Exp Psychol: General 146(12):1705-1721, 2017, keyworded as an adversarial collaboration) paired masked letter strings with words from semantic categories. They found unconscious conditioning with a brief (75 ms) CS-US interval and an effectively masked CS, with no reportable knowledge of the contingency. With a longer (375 ms) interval and a visible CS, conditioning tracked awareness and did not occur without it. Some learning therefore happens without awareness, but it differs from conscious conditioning.

    2017 · faculty.washington.edu

  3. With masking, a cue captured attention whether or not people saw it. Without awareness, attention was less stable over time and inhibition of return was reduced. Attention survives without awareness, but its control changes. This partly fits the adaptation view.

    2016-06 · collaborate.princeton.edu

  4. Bronfman, Ginsburg and Jablonka, 'The transition to minimal consciousness through the evolution of associative learning' (Frontiers in Psychology 7:1954, 2016). They propose UAL as the marker of the evolutionary transition to minimal consciousness and as a driver of its evolution, and they offer testable predictions. This is a theoretical proposal, not an experiment.

    2016 · cris.openu.ac.il

  5. Lovibond, Liu, Weidemann and Mitchell (Biological Psychology 87(3):393-400, 2011) used the Clark and Squire procedure in two experiments. Both trace and delay eyeblink procedures produced differential conditioning only in participants who showed contingency awareness, measured by a post-experiment questionnaire. This contradicts the claim that delay conditioning runs without awareness. The authors argue against separate aware and unaware learning systems.

    2011 · researchers.westernsydney.edu.au

  6. Balderston and Helmstetter (Behavioral Neuroscience 124(4):478-489, 2010) conditioned participants on backward-masked shapes that they could not reliably identify. The training changed later skin conductance responses but not their expectancy ratings for the shock, which the authors read as learning without awareness under a dual-process model. They cite Knight and colleagues' finding that awareness is not needed for fear expression in delay conditioning but is needed in trace conditioning. They also cite Esteves et al. (1994), whose masked trace conditioning with fear-relevant stimuli produced conditioning. So the dissociation between delay and trace conditioning comes from Knight et al., which this paper cites but did not test.

    2010 · pmc.ncbi.nlm.nih.gov

The theories

Grouped by what each one takes as the starting point. Open any theory for its claims, predictions, objections and recent changes, each with its source.

Experience comes first

Consciousness, or its structure, is basic. Matter and spacetime are built from it or alongside it.

Conscious realismEvolution tuned perception to fitness, not truth, so space, time and objects are a user interface; the underlying reality is a network of conscious agents, described by Markov dynamics.

Conscious realism, Interface Theory of Perception, Fitness-Beats-Truth, conscious agents and trace logic. Donald Hoffman, Chetan Prakash, Manish Singh, Chris Fields, Robert Prentner, Justin Mark, Brian Marion. Last checked 15 Nov 2026.

Core claims

  • Fitness beats truth: in evolutionary game simulations, perceptual strategies tuned to payoffs can drive veridical strategies to extinction, so natural selection need not favor accurate perception (Mark, Marion and Hoffman, Journal of Theoretical Biology 2010). The result was later generalized as the Fitness-Beats-Truth theorem (Prakash, Stephens, Hoffman, Singh and Fields, Acta Biotheoretica 2021). pubmed.ncbi.nlm.nih.gov
  • Interface Theory of Perception: perception is an adaptive interface, like a desktop, that hides objective structure and shows fitness-relevant icons. Hoffman, Singh and Prakash state it in Psychonomic Bulletin & Review (2015), alongside commentaries and a reply. featuredcontent.psychonomic.org
  • Conscious realism with a formal model of a 'conscious agent': the world consists of interacting conscious agents, and physical objects and spacetime may be patterns in their dynamics. The paper rejects the idea that objects exist unperceived and the claim that brains cause consciousness (Hoffman and Prakash, Frontiers in Psychology 2014). pmc.ncbi.nlm.nih.gov
  • Spacetime is not fundamental. In 'Fusions of Consciousness' (Entropy 2023), agent dynamics are Markov chains. The possible dynamics of n agents form a Markov polytope and their possible fusions a fusion simplex. A map from Markov chains to decorated permutations links agent dynamics to the structures physicists use for scattering amplitudes, so spacetime and scattering are a data structure coding agent interactions, and a particle is a projection of a communicating class of agents. pmc.ncbi.nlm.nih.gov
  • Trace logic (2024 preprint, 'Traces of Consciousness'): a trace order on Markov chains induces a non-Boolean, locally Boolean logic on Markov kernels. One agent observes another when its kernel sits below the other's in the trace order, and there is a homomorphism to the Lebesgue logic of probabilistic beliefs. preprints.org
  • Prentner and Hoffman (2024) apply conscious agent theory and the interface theory to consciousness science and AI. Computation describes the dynamics of consciousness as they appear on the interface. Since consciousness already exists, building AI does not create it, though AI may help us understand it. And we live in a simulation instantiated in consciousness, not on a digital substrate. frontiersin.org

How it explains the evolution of consciousness

Biologically, selection shapes perceptual interfaces for fitness payoffs rather than truth, so organisms see species-specific icons, not objective structure (Mark et al. 2010; FBT theorem 2021). Cosmically, Hoffman takes consciousness as primary, not as something that evolved from matter. Agents exist outside spacetime and combine into more complex agents or fuse into new ones, which is his formal take on the combination problem. Spacetime, bodies and brains are the interface through which agents appear. Prentner and Hoffman (2024) add that treating agency as exclusive to living beings may hold only for practical purposes, not as a fundamental truth. Darwinian evolution is therefore a description inside the interface. Hoffman wants to recover it as a projection of the deeper agent dynamics, but has not derived it yet.

Predictions and where they stand

  • Mixed

    Evolutionary games: strategies tuned to fitness payoffs outcompete strategies that estimate the true world state (Fitness-Beats-Truth).

    Shown in simulations (2010) and proven as a theorem under its stated assumptions (2021). How far it generalizes is disputed. Berke, Walter-Terrill, Jara-Ettinger and Scholl (Cognitive Science 2022) found that when one inflexible perceptual system must serve several independent goals, evolution increasingly favors veridical representations. Critics also argue that fitness is correlated with truth in realistic settings. api.philpapers.org

  • Untested

    Perceived objects lack observer-independent definite properties and do not exist unperceived. Perception is a species-specific interface and need not resemble the underlying reality.

    Stated in Hoffman and Prakash 2014 and Hoffman, Singh and Prakash 2015, with an appeal to quantum results on unobserved properties. Commentators on the 2015 target article disputed it, among them Lupyan, who argued perception is 'true enough'. No experiment has been designed to decide between this view and ordinary accounts of perception. featuredcontent.psychonomic.org

  • Not yet testable

    Spacetime and particles are projections of Markov dynamics among conscious agents, linked via decorated permutations to the positive-geometry structures physicists use for scattering amplitudes.

    Fusions of Consciousness defines the map from Markov chains to decorated permutations. No derivation of known scattering amplitudes, and no new quantitative physical prediction, has been published. pmc.ncbi.nlm.nih.gov

  • Untested

    Observation between conscious agents is a trace order on Markov kernels, and the resulting logic maps homomorphically to the Lebesgue logic of probabilistic belief.

    This is a mathematical claim in a 2024 preprint. It is not an empirical prediction, and no independent peer review or verification of it has been located. preprints.org

Strongest objections

  • Berke, Walter-Terrill, Jara-Ettinger, Scholl (Cognitive Science 2022). Earlier evolutionary models assumed perception serves a single goal. When one automatic, inflexible perceptual system must serve many independent goals, simulations favor veridical representations, so fitness-optimal perception can be accurate. api.philpapers.org
  • Gary Lupyan (Psychonomic Society featured content, 2015). Evolutionary games do not show perception is nonveridical. Fitness-only models are too narrow, the theory confuses perception with action and categorization, and perception is 'true enough' to support discoveries like measuring the speed of light. featuredcontent.psychonomic.org
  • Jonathan Cohen (Psychonomic Bulletin & Review 2015 commentary, 'Perceptual representation, veridicality, and the interface theory of perception'). The arguments against standard veridical perception are not persuasive, and the interface theory itself seems committed to veridical representation. featuredcontent.psychonomic.org
  • 4gravitons (blog by a former theoretical particle physicist), review of The Case Against Reality, 2024. FBT overlooks that fitness is correlated with truth. Hoffman misreads physicists: 'spacetime is doomed' is borrowed from Arkani-Hamed without his reasons. His realist model, with real observers communicating in a definite order, contradicts the non-realist quantum interpretations he cites. Open questions include whether the agent networks are Turing-complete, how they fit relativity and quantum mechanics, and how evolution applies if agents can be created or destroyed. 4gravitons.com

How it has been changing

  1. 2024-10The preprint 'Traces of Consciousness' introduces the trace order and trace logic on Markov kernels. No peer-reviewed version has been located. preprints.org
  2. 2024-07-15Prentner and Hoffman publish 'Interfacing consciousness' (Perspective, Frontiers in Psychology). It extends the theory to AI, treating computation as an interface-level description, and proposes that we live in a simulation instantiated in consciousness. It states no specific falsifiable test. frontiersin.org
  3. 2024-02-23A physics-oriented book review argues that Hoffman's appeal to 'spacetime is doomed' and his realist agent model do not fit the physics he cites. 4gravitons.com
  4. 2023-01-09Hoffman, Prakash and Prentner publish 'Fusions of Consciousness' in Entropy. It adds the Markov polytope, the fusion simplex and a map from Markov chains to decorated permutations, moving the program toward scattering-amplitude physics. pmc.ncbi.nlm.nih.gov
  5. 2022Berke et al. publish evolutionary simulations in Cognitive Science showing that with multiple independent goals, evolution favors veridical perception, the most prominent peer-reviewed challenge to how far FBT generalizes. api.philpapers.org

Open questions inside the theory

  • Can any experiment distinguish the interface theory from ordinary evolutionary perception science, given that accuracy appears favored when perception serves many goals?
  • Does the decorated-permutation or positive-geometry link actually yield known scattering amplitudes, or a new quantitative physical prediction?
  • How does fusing agents yield a unified experience (the combination problem)?
  • Will independent mathematicians confirm the trace logic results, and what do they add beyond definitions?
  • How does Darwinian evolution inside the interface map onto the deeper agent dynamics?
  • Is the position coherent idealism, or does it quietly assume a veridical objective structure (the agent network, observers in a definite order) at odds with its own anti-realism?
Integrated InformationExperience is identical to a physical system's irreducible cause-effect structure; the more a system's parts constrain each other as a unified whole (integrated information, phi), the more it is conscious.

Integrated Information Theory (IIT 4.0). Giulio Tononi, Christof Koch, Larissa Albantakis, William Marshall, Graham Findlay, Melanie Boly, Marcello Massimini. Last checked 15 Nov 2026.

Core claims

  • IIT 4.0 starts from axioms about experience, derives physical postulates (intrinsicality, information, integration, exclusion, composition) and expresses them mathematically. It introduces a unique measure of intrinsic information consistent with the postulates and an explicit assessment of causal relations. pmc.ncbi.nlm.nih.gov
  • By unfolding a system's full irreducible cause-effect power, the distinctions and relations it specifies are claimed to account for the quality of experience, not only its presence. The theory claims testable predictions about both. pmc.ncbi.nlm.nih.gov
  • Consciousness is graded, is common among biological organisms and can occur in some very simple systems. Purely feed-forward networks, however complex, are predicted not to be conscious, and neither are aggregates such as groups of individuals or heaps of sand. pmc.ncbi.nlm.nih.gov
  • Digital computers would experience next to nothing even if they simulated a human brain, because what matters is the physical causal structure of the hardware, not the software it runs. pmc.ncbi.nlm.nih.gov
  • Proponents locate the main neural substrate of experience in humans in a posterior cortical 'hot zone' rather than a fronto-parietal network. In IIT 4.0 terms, that substrate is the set of units that forms a maximum of intrinsic cause-effect power. nature.com

How it explains the evolution of consciousness

IIT has no evolutionary story in the standard sense. Because consciousness is identified with intrinsic cause-effect structure, it is a property of any physical system with the right causal organization, not an adaptation that appeared at a particular point. Per Tononi and Koch (2015), it is graded, widespread among biological organisms and possible in some very simple systems. Evolution matters as a process that builds substrates with highly integrated structure, such as densely recurrent cortical circuits. Proponents distinguish IIT from panpsychism: the exclusion postulate denies consciousness to aggregates and to feed-forward systems. Barrett et al. (2026) nonetheless argue that the theory implies space (and time) are tiled with substrates of (proto-)consciousness, and that this is not a problem for it. No source was found in which IIT authors give a worked account of the cosmic or evolutionary origin of consciousness, so none is asserted here.

Predictions and where they stand

  • Supported

    Conscious content is decodable from posterior cortex, and adding prefrontal activity does not improve decoding.

    In the Cogitate adversarial collaboration, category and orientation could be decoded from posterior areas, and adding prefrontal regions did not improve decoding and sometimes reduced it. The preregistration treated this as a non-critical test of IIT. pmc.ncbi.nlm.nih.gov

  • Mixed

    Conscious content should be maintained in posterior cortex for as long as a stimulus is perceived.

    IIT passed the predefined criteria for the duration prediction: some posterior electrodes, mostly early visual and ventral temporal, tracked stimulus duration. Tracking was sparse, though (about 15% of face-selective electrodes), and no sustained representation of orientation was found. The authors treat the latter as a challenge for IIT. pmc.ncbi.nlm.nih.gov

  • Challenged

    Conscious perception should involve sustained synchronization (gamma-band connectivity) within posterior cortex.

    Intracranial and MEG data showed brief, mostly low-frequency synchrony, largely explained by the evoked response. Based on the preregistration, the authors call the lack of sustained posterior synchronization the most direct challenge to IIT. pmc.ncbi.nlm.nih.gov

  • Supported

    The brain's response to a perturbation (TMS-EEG) is more integrated and differentiated when conscious, so perturbational complexity (PCI) drops in dreamless sleep and anaesthesia.

    PCI is lower in deep sleep and anaesthesia and has clinical use in disorders of consciousness. Barrett et al. note the results are equally compatible with weaker versions of IIT and possibly with global workspace theory, and that PCI is a proxy for phi, not an approximation with a bounded error. arxiv.org

  • Untested

    Inactive neurons that could fire still contribute to experience. Optogenetically silencing them, so they can no longer fire, should change conscious perception even though their activity does not change.

    The INTREPID collaboration, funded by the Templeton World Charity Foundation, pits IIT against two predictive processing accounts and includes an optogenetic silencing experiment in mice, with each experiment to be replicated in a second lab. A February 2025 TWCF post describes it as ongoing and multi-year. No results were reported. templetonworldcharity.org

  • Not yet testable

    Grid-like connectivity underlies the extendedness of spatial experience, and directed-chain connectivity underlies the felt flow of time.

    Barrett et al. describe this as part of IIT's account of experiential quality, with colour and conceptual hierarchy still in progress. Phi has not been computed for any real physical system, so direct tests are not yet possible. arxiv.org

Strongest objections

  • Stephen Fleming, Hakwan Lau and others (open letter, 124 signatories). The letter, posted on PsyArXiv on 15 September 2023, argued that IIT should be labelled pseudoscience. It cited its panpsychist implications, the doubtful testability of its core assumptions, and the claim that recent experiments did not test those core ideas. Anil Seth called the label inflammatory. Liad Mudrik, a co-lead of the adversarial study, defended IIT's testability, and Erik Hoel said the letter singled IIT out for problems other theories share. scientificamerican.com
  • IIT-Concerned (Klincewicz, Cheng, Schmitz, Sebastian, Snyder and co-signatories), Nature Neuroscience 2025. The comment argues that IIT's core claims cannot be tested even in principle, so the theory is unscientific. It separates those core claims from the empirical results proponents present as partial support. pubmed.ncbi.nlm.nih.gov
  • Scott Aaronson. IIT implies that a large, suitably structured network of XOR gates (an expander) would be more conscious than a human. Tononi accepted this consequence. Aaronson objected not to panpsychism as such but to what he saw as an arbitrary and selective attribution of consciousness. scottaaronson.blog
  • Adam Barrett, Anil Seth, Pedro Mediano and colleagues (2026 review). Phi is not well defined for real physical systems: the algorithm requires Markovian dynamics, and its uniform prior fails for continuous variables. Phi has not been computed on any real system, and proxies such as PCI are not approximations with known error. Fitting current physics would require a reformulation in terms of continuous fields. The authors also defend IIT against common misreadings, for example that it gives substantial consciousness to an inactive grid of logic gates. arxiv.org
  • Cogitate Consortium results. The Nature paper reports that the data align with some predictions of IIT and of global neuronal workspace theory while substantially challenging key tenets of both. For IIT, the main failure was the absence of sustained synchronization within posterior cortex. pubmed.ncbi.nlm.nih.gov

How it has been changing

  1. 2026-04Barrett, Milinkovic, Mediano, Rosas, Bor, Barnett and Seth posted 'Integrated information theory: the good, the bad and the misunderstood' on arXiv. It proposes a multi-dimensional characterisation of global conscious states and a 'weak' IIT focused on explanatory correlates, and stresses unsolved technical problems. arxiv.org
  2. 2025-11Marshall, Findlay, Albantakis and Tononi released 'Evaluating Artificial Consciousness through Integrated Information Theory', which applies IIT to artificial systems as an alternative to computational functionalism. Its publication venue is unconfirmed; the PDF is hosted by IONS. noetic.org
  3. 2025-04-30Cogitate results appeared in Nature (vol. 642, pp. 133-142). Some IIT predictions held, the sustained posterior synchronization prediction failed, and global workspace predictions about the prefrontal cortex also fared poorly. pubmed.ncbi.nlm.nih.gov
  4. 2025-03IIT-Concerned published the Nature Neuroscience comment 'What makes a theory of consciousness unscientific?'. Tononi and colleagues published a reply in the same journal. pubmed.ncbi.nlm.nih.gov
  5. 2023-09-15124 scholars posted an open letter calling IIT pseudoscience, setting off a public dispute over the label and its effect on the field. nature.com
  6. 2023-06The Cogitate preregistered adversarial collaboration (IIT vs global neuronal workspace) released its results as a bioRxiv preprint before the journal version. biorxiv.org
  7. 2022-12 (preprint); 2023-10 (journal)IIT 4.0 was published in PLOS Computational Biology (arXiv preprint 2212.14787). It adds a new intrinsic information measure, an explicit treatment of causal relations and a tighter mapping from axioms to postulates. pmc.ncbi.nlm.nih.gov
  8. 2017-04-10High-density EEG showed that dreaming in both REM and NREM sleep tracks activity in a posterior cortical hot zone, and that activity there correlated with specific dream contents. This fits the posterior focus of IIT. doi.org

Open questions inside the theory

  • Can phi, or a principled approximation with known error, be computed for real brains, given non-Markovian dynamics and continuous variables?
  • Does IIT make any core prediction, as opposed to correlational ones like PCI, that can separate it from rival theories?
  • Do inactive neurons contribute to experience, as IIT says? INTREPID is the designed test, and no results have been published.
  • Should the field adopt a 'weak' IIT of explanatory correlates, or does the 'strong' identity between experience and cause-effect structure hold?
  • How can IIT be reconciled with continuous-field physics?
  • Does the claim that conventional digital hardware cannot be conscious hold up as AI systems scale, and could neuromorphic hardware change the answer?
PanpsychismPhysics tells us what matter does, not what it is. Perhaps its hidden nature is experience, so consciousness is basic and not produced from nothing by brains.

Panpsychism / Russellian monism. Philip Goff, Galen Strawson, David Chalmers, Hedda Hassel Mørch, Luke Roelofs, Sam Coleman, Angela Mendelovici, William Seager, Bertrand Russell, Arthur Eddington. Last checked 15 Nov 2026.

Core claims

  • Panpsychism is 'the doctrine that mind is a fundamental feature of the world which exists throughout the universe'. Its basic rival is emergentism, on which mind appears only at certain times and places (Seager, original 2001 SEP entry, archived version). plato.stanford.edu
  • Russellian monism combines three theses: structuralism about physics (physics describes only spatiotemporal structure and dynamics), realism about quiddities (properties underlie that structure), and quidditism about consciousness (quiddities are relevant to phenomenal consciousness). In its panpsychist form quiddities are themselves phenomenal; in its panprotopsychist form they are protophenomenal and collectively constitute phenomenal properties. The SEP entry (Alter and Pereboom, revised 2023) says there is no consensus on exactly how quiddities are relevant to consciousness. plato.stanford.edu
  • The current SEP panpsychism entry (substantive revision 13 May 2022) gives two main arguments: the anti-emergence argument (consciousness cannot intelligibly arise from wholly non-conscious matter) and the intrinsic nature argument (physics reveals only structure, and the one intrinsic nature we know directly, in our brains, is conscious). Among other arguments it lists Clifford's genetic argument from evolutionary continuity. plato.stanford.edu
  • Strawson (2006) argues that a realistic physicalist must accept panpsychism: consciousness is real and cannot emerge from matter as ordinarily conceived, so the matter in brains must have an experiential nature. The mystery is what matter is, not what consciousness is. informationphilosopher.com
  • Chalmers (in Brüntrup and Jaskolla, eds, Panpsychism, OUP) splits the combination problem into subject, quality and structure combination problems: how micro-level experiences could combine into macro-level ones. He says he has neither solved it nor proved it unsolvable, and surveys options with their costs. consc.net
  • Goff defends panpsychism by inference to the best explanation, not experiment. On 3 December 2025 he wrote that 'You can't do experiments to test whether an electron or a field has feelings', and noted that people never demand to know how we test physicalism either. philipgoff.substack.com

How it explains the evolution of consciousness

Continuity, not a late jump. On the genetic or continuity argument (Clifford's, cited in the SEP), evolution reshapes and combines what already exists and does not create a wholly new kind of property, so experience must have roots in the simplest matter and organisms; Clifford held that even in the amoeba there is something 'of the same nature with our own consciousness'. Human consciousness is then a complex combination or organisation of simpler experience (constitutive panpsychism), or something the micro-level gives rise to (emergentist panpsychism). Cosmopsychism reverses the direction: the universe as a whole is the basic subject and individual minds derive from it. Goff's 2023 book Why? The Purpose of the Universe argues for cosmic purpose and gives sustained attention to a conscious universe among the candidate explanations. Panpsychism offers no mechanism for when human-level minds arose, only that experience was never wholly absent.

Predictions and where they stand

  • Not yet testable

    Panpsychism as a metaphysical thesis makes no distinctive experimental prediction; it competes with physicalism and dualism on explanatory grounds.

    Goff, 3 December 2025: 'You can't do experiments to test whether an electron or a field has feelings.' He holds that any account of why physical activity goes with experience will rest on inference to the best explanation. philipgoff.substack.com

  • Challenged

    IIT (a theory whose proponents say consciousness is a fundamental feature of the universe, which critics read as panpsychist) predicts sustained synchronization within posterior cortex during conscious perception.

    The Cogitate adversarial collaboration (Nature, published online 30 April 2025; 256 participants; fMRI, MEG and intracranial EEG) found no sustained posterior synchronization, which challenges IIT. It also challenged GNWT (little ignition at stimulus offset, limited prefrontal representation). The results partly matched predictions of both theories. This tests IIT, not panpsychism as such. ebi.ac.uk

  • Not yet testable

    Would panpsychism make large language models conscious? Chalmers says assuming everything is conscious gives 'a very easy road' to LLM consciousness. Most panpsychists, though, attribute experience to fundamental entities and deny that every aggregate, such as a computer, is a unified subject, so the view does not by itself imply that LLMs are subjects.

    Chalmers explicitly does not assume panpsychism in his analysis and puts his credence in current LLM consciousness 'somewhere under 10 percent' (Boston Review, 9 August 2023, an edited version of his NeurIPS talk of 28 November 2022). arxiv.org

Strongest objections

  • Sam Coleman. Subjects' points of view exclude one another, so micro-subjects cannot sum into a macro-subject. He argues that constitutive panprotopsychism handles combination better than panpsychism and defends 'panqualityism', with qualities but no micro-subjects. doi.org
  • David Chalmers. The combination problem, especially the subject-summing problem, is the central unsolved difficulty, and a solution has to address the subject, quality and structure versions together. He surveys options without claiming a solution. consc.net
  • Angela Mendelovici. Argues that panpsychism's combination problem is a problem for everyone: rival views face analogous problems, so it does not by itself favour them. philarchive.org
  • Massimo Pigliucci and Sean Carroll (as reported by Jerry Coyne, Why Evolution Is True, 2021). Pigliucci: panpsychism 'doesn't make any contact with the empirical world' and replaces one mystery with another. Carroll: attributing a new property to particles conflicts with the Standard Model. whyevolutionistrue.com
  • Commentators on Strawson's 'Realistic Monism' (McGinn, Coleman, Rey, Smart), per Leopold Stubenberg's 2007 NDPR review. Strawson's view may not be physicalism at all; it can look like 'dualism all the way down', which undercuts the claim that physicalism entails panpsychism. ndpr.nd.edu
  • Objections catalogued in the SEP entry (2022). The incredulous stare (the view seems absurd), the combination and subject-summing problems, the palette problem (a few micro-qualities must yield the vast range of experienced qualities), and the structural mismatch problem, whose most-discussed form is the grain problem. plato.stanford.edu

How it has been changing

  1. 2025-12Goff states plainly that panpsychism cannot be tested by experiment and rests on inference to the best explanation. philipgoff.substack.com
  2. 2025-04The Cogitate adversarial collaboration (Nature, online 30 April 2025) reported results that challenge key tenets of IIT, the theory most often linked with panpsychism, and of GNWT. The metaphysical thesis of panpsychism is not itself tested. ebi.ac.uk
  3. 2025In the Nature Neuroscience comment 'What makes a theory of consciousness unscientific?' (Klincewicz, Cheng, Schmitz, Sebastián, Snyder and IIT-Concerned; vol. 28, 689-693), the letter's authors argue that IIT's core claims are untestable even in principle. They single out its panpsychist-leaning claims that 'the entire cosmos is suffused with sentience' and that inactive grids of logic gates are conscious. uwo.ca
  4. 2024Goff's Journal of Consciousness Studies paper 'How Exactly Does Panpsychism Help Explain Consciousness?' (vol. 31, no. 3) argues that strong and weak emergentist panpsychism each inherit problems from dualism or physicalism. He proposes a hybrid: subjects are strongly emergent and their phenomenal properties weakly emergent. citedrive.com
  5. 2023-09An open letter signed by more than 100 researchers called IIT pseudoscience, chiefly over its testability. Critics of the letter, such as Erik Hoel, said the signatories were 'acting irresponsibly'. It is the main public fight over panpsychism-adjacent consciousness science. psychologytoday.com
  6. 2023-09Goff's Why? The Purpose of the Universe (Oxford UP; online 21 September 2023, print 9 November 2023) rejects the 'Omni-God', defends cosmic purpose and weighs alternatives, among them a limited designer, a simulation, purposive laws and a universe that is itself a conscious mind. academic.oup.com

Open questions inside the theory

  • Is there any version of panpsychism with a distinctive, testable empirical consequence, or does it stay a choice made on explanatory grounds?
  • Can the subject-summing problem be solved by phenomenal bonding, fusion, deflationism about subjects or cosmopsychism, or does it push toward panprotopsychism or panqualityism?
  • Do failures of IIT predictions bear on panpsychism at all, or are the two logically independent?
  • Micro or cosmic: should the fundamental subject be particles or the universe as a whole?
  • What separates aggregates that are unified subjects (brains) from those that are not (rocks, computers), and does the answer settle whether AI systems could be conscious?

What the brain does

Consciousness is a kind of information processing in brains: broadcast, monitoring, or a model that seems ineffable.

Global WorkspaceMany unconscious brain processors run in parallel. A stimulus becomes conscious when it wins a competition, 'ignites' a large fronto-parietal network, and is broadcast for flexible use and report.

Global (Neuronal) Workspace Theory. Bernard Baars, Stanislas Dehaene, Jean-Pierre Changeux, Lionel Naccache, George Mashour, Pieter Roelfsema. Last checked 15 Nov 2026.

Core claims

  • Baars proposed global workspace theory in 1988 (A Cognitive Theory of Consciousness) and extended it in 1997 (In the Theater of Consciousness). Specialised unconscious processors compete, and the winning content is broadcast to memory, language and motor planning. Baars rejects a homunculus watching the stage. en.wikipedia.org
  • Dehaene, Kerszberg and Changeux gave the idea a neural form in 1998 (PNAS): a workspace of long-axon neurons linking distant cortical areas. ugr.es
  • Dehaene and Naccache (2001) argue that many modular networks work unconsciously in parallel, and that global availability of information is what is subjectively experienced as a conscious state. home.cs.colorado.edu
  • Dehaene and Changeux (2011, Neuron) tie conscious access to long-distance beta/gamma synchronization and 'ignition' of a prefronto-parietal network. They also discuss anesthesia, coma and vegetative states. unicog.org
  • Mashour, Roelfsema, Changeux and Dehaene (2020, Neuron) define ignition as sudden, coherent and exclusive activation of workspace neurons coding the current content, with the rest inhibited. They give large layer II/III pyramidal cells a key role and link conscious access to attention and working memory. pubmed.ncbi.nlm.nih.gov
  • Dehaene, Lau and Kouider (2017, Science) split consciousness in two. C1 is selection of information for global broadcast. C2 is self-monitoring that produces a sense of certainty or error. They argue that current machines mostly carry out unconscious (C0) computation. science.org

How it explains the evolution of consciousness

Consciousness is treated as an evolved architecture, not a cosmic fundamental. The workspace evolved to break the brain's modularity, so that specialised processors can share information and flexibly combine to solve novel tasks. Dehaene and Naccache (2026) state that this holds 'in humans and other animals'. On this view a functionally equivalent architecture could in principle arise in machines. Zacks and Jablonka (2023, Neuroscience of Consciousness) trace the workspace's evolutionary origins in vertebrates. They argue that in basal vertebrates the event memory system in the hippocampal homolog plays the workspace role, and that GNW needs important modifications to apply there.

Predictions and where they stand

  • Mixed

    Conscious access is all-or-none, marked by a late non-linear 'ignition' with a P3b-like signal, while unconscious processing stays graded and local.

    In the attentional blink, visibility ratings and P3 responses look all-or-none (Sergent and Dehaene 2004; Sergent, Baillet and Dehaene 2005). Bowman and colleagues add EEG support and explain the pattern by delayed consolidation. They also note that masking studies show a more continuous, though non-linear, gradation of visibility. cs.kent.ac.uk

  • Challenged

    The P3b is a signature of conscious perception, not just of reporting.

    Cohen, Ortego, Kyroudis and Pitts (J Neurosci 2020) used visual masking. They found a P3b for visible stimuli only when the stimuli were task-relevant and reported, and none in the no-report condition. They concluded that the P3b reflects post-perceptual processing and is not a neural signature of consciousness, while noting that this does not disprove cognitive theories. pmc.ncbi.nlm.nih.gov

  • Supported

    Prefrontal cortex carries the contents of conscious perception, not only report-related signals.

    Kapoor et al. (Nature Communications 2022) recorded from two macaques in binocular rivalry without report. Feature-selective ventrolateral prefrontal neurons tracked perceptual changes, and single-trial population decoding of conscious content worked, including for internally generated transitions. This shows prefrontal involvement, not ignition itself. pmc.ncbi.nlm.nih.gov

  • Challenged

    Prefrontal ignition should appear at both the onset and the offset of a conscious stimulus, and prefrontal cortex should represent the conscious dimensions of the stimulus.

    The preregistered Cogitate adversarial test (256 participants; fMRI, MEG, iEEG) found some conscious information in prefrontal cortex. The authors name two main challenges to GNWT: a general lack of ignition at stimulus offset, and limited prefrontal representation of some conscious dimensions. Neither theory is declared refuted. biorxiv.org

  • Mixed

    Conscious access involves long-range, content-specific synchronization between prefrontal cortex and the sensory areas coding the content.

    Cogitate found content-specific synchronization between frontal and early visual areas, which partly fits GNWT. Sustained occipital and lateral temporal activity tracking stimulus duration was an IIT-favoured result, not GNWT support. Press coverage says the results matched some predictions of both theories. healthenews.mcgill.ca

  • Untested

    Because GNW is a functional architecture, a workspace-like broadcast stage with global availability (C1) and self-monitoring (C2) could be found in artificial systems and tested there.

    Dehaene and Naccache's June 2026 commentary calls a reported workspace-like band of intermediate layers (the 'J-space') in Claude Sonnet 4.5 a landmark, because it gives a mechanistic, testable version of GNW. They conclude that the model approximates the functional architecture of conscious processing. They also stress key differences: its anatomy, its sense of self, and its lack of a body and of an enduring episodic memory, which warrant caution in drawing parallels with the human mind. They propose further tests. unicog.org

Strongest objections

  • Ned Block and overflow defenders. Block argues that perceptual experience is richer than what can be cognitively accessed (Sperling-type partial report), so access cannot be the whole of phenomenal consciousness. Naccache and Dehaene reply in BBS that the apparent richness is an illusion of phenomenality that the workspace model can explain. cambridge.org
  • Michael Pitts and colleagues. No-report paradigms suggest that much of what looked like the neural signature of consciousness (P3b, late frontal activity) reflects reporting and post-perceptual processing. pmc.ncbi.nlm.nih.gov
  • Cogitate Consortium (including Melloni, Mudrik, Koch). The 2025 preregistered test did not confirm offset ignition and found limited prefrontal representation of some conscious dimensions. GNWT was challenged on key tenets, not refuted. idw-online.de

How it has been changing

  1. 2026-06Dehaene and Naccache published a commentary on an Anthropic paper (Gurnee et al., 'Verbalizable Representations Form a Global Workspace in Language Models') that reports a workspace-like 'J-space' in Claude Sonnet 4.5. They call it a landmark and a mechanistic, testable form of GNW. They judge that the model approximates the functional architecture of conscious processing but differs in key ways: anatomy, sense of self, no body, and no enduring episodic memory. unicog.org
  2. 2026-02Lecture 4 of Dehaene's 2025-2026 Collège de France course ('Distilling the neural correlates of conscious perception') answers no-report challenges to prefrontal cortex and the P3b. For example, it argues that weaker frontal activity under no-report may reflect limited sensitivity, and that those signals were tied to transitions, not content. These are lecture slides, not peer-reviewed work. college-de-france.fr
  3. 2025-04-30Cogitate published in Nature (vol 642, pp 133-142). The results match some predictions of both theories and substantially challenge key tenets of both. Data are public, and a second large experiment is reported as underway. healthenews.mcgill.ca
  4. 2023-09An open letter from 124 researchers called IIT pseudoscience and argued that the Cogitate experiment did not test IIT's core claims. The dispute centred on IIT and on how the adversarial test was framed. carlzimmer.com
  5. 2023-06Cogitate preprint posted on bioRxiv. The abstract names lack of ignition at stimulus offset and limited prefrontal representation of some conscious dimensions as the main challenges to GNWT. biorxiv.org
  6. 2023Zacks and Jablonka (Neuroscience of Consciousness) argue that in basal vertebrates the workspace is instantiated by the event memory system in the hippocampal homolog, and that GNW needs important modifications there. researchonline.lse.ac.uk
  7. 2022-03-22Macaque prefrontal neurons tracked perception during a no-report task with internally generated switches. Single-trial population decoding of conscious contents worked without volitional reports or changes in sensory input, supporting prefrontal content coding. doi.org
  8. 2021-05Melloni, Mudrik, Pitts and Koch set out the adversarial collaboration method in Science: find divergent predictions, preregister them, then test. Cogitate, funded by Templeton, pits GNWT against IIT. christofkoch.com
  9. 2015-10-05When perceptual switches were made unreportable, fronto-parietal BOLD responses to them almost vanished. The authors read frontal activity as a consequence of switches, not their cause, which presses on prefrontal accounts. doi.org

Open questions inside the theory

  • If prefrontal ignition is weak or absent at stimulus offset, does GNWT drop that prediction or revise what ignition tracks?
  • Can no-report paradigms separate consciousness from access, or do they lose sensitivity to real frontal signals?
  • Is access the whole story, or does phenomenal experience overflow the workspace, as Block argues?
  • Does the workspace need recurrent loops over time, or can a single forward pass in a language model count?
  • What will the second Cogitate experiment and the adversarial collaboration with higher-order theories show? I found no published results for the latter.
  • How did workspace architecture arise across vertebrate lineages, and is it present in animals without a cortex?
Higher-OrderA mental state is conscious when the mind represents or monitors it in a higher-order way. Variants differ on whether that is a thought, an inner perception, or a subpersonal reliability tag.

Higher-Order Theories (incl. Perceptual Reality Monitoring). David Rosenthal, Hakwan Lau, Richard Brown, Stephen Fleming, Axel Cleeremans, Matthias Michel, Joseph LeDoux, Peter Carruthers, William Lycan. Last checked 15 Nov 2026.

Core claims

  • The broad family: a mental state is phenomenally conscious when it is the object of a higher-order representation of the right kind. Versions include higher-order perception (inner sense), actualist HOT (Rosenthal), dispositionalist HOT (Carruthers), and self-representational and hybrid views. plato.stanford.edu
  • Rosenthal's actualist view: a state is conscious when it is accompanied by a suitable non-inferential higher-order thought that one is in that state. Both Rosenthal and Lycan allow that a targetless higher-order representation can occur, rarely, and hold that the subject is then phenomenally conscious. Lycan's example is morphine patients; Rosenthal's is dental patients whose pain seems real. plato.stanford.edu
  • Perceptual Reality Monitoring (PRM), developed by Lau (In Consciousness We Trust, 2022) and set out with Michel: a sensory state is conscious when a subpersonal higher-order mechanism, plausibly prefrontal, tags it as a reliable reflection of the present world rather than noise or imagination. The tag concerns reliability; it is not a copy of the content and not personal-level metacognition. philarchive.org
  • Brown, Lau and LeDoux (2019) say critics misread the approach. Higher-order states demand far less cognition than often assumed. Most proponents do not equate consciousness with metacognition or confidence. The view sits between global workspace accounts and early-sensory (local recurrence) accounts. philarchive.org
  • Brown argues that the evidence for phenomenal overflow (Block) does not establish overflow and fits a higher-order account. He later developed HOROR (Higher-Order Representation of a Representation), a non-relational variant on which the higher-order representation is itself the phenomenally conscious state. philarchive.org
  • ETHOS, the Templeton-funded project led by Fleming and Cleeremans, frames the family's internal disputes along two axes. The first is rich versus sparse higher-order content: does it duplicate the first-order content or only track its strength or reliability? The second is non-relational versus relational: can the higher-order state misrepresent the first-order state or not? templetonworldcharity.org

How it explains the evolution of consciousness

Not a theory of origins in itself. Higher-order theories imply that consciousness needs some self-monitoring or meta-representational machinery, so it would arise wherever a system evolves, or is built, with that machinery. The SEP entry names the main evolutionary pressure point: actualist and dispositionalist HOT seem to require mental-state concepts that many animals may lack. Defenders answer with simpler or implicit higher-order representations and with evidence of metacognition in other species. Brown, Lau and LeDoux argue the demands are lower than assumed. PRM ties the monitoring mechanism to mammalian prefrontal cortex, which narrows the likely range of conscious creatures. The same logic is applied to AI. Butlin, Long, Chalmers et al. (2023) use higher-order indicators, such as metacognitive monitoring that separates reliable perceptual representations from noise, to assess machines. They conclude that no current system is a strong candidate, but that no obvious technical barrier stands in the way. I found no cosmic or deep-time evolutionary account in this literature.

Predictions and where they stand

  • Mixed

    Prefrontal activity should track conscious perception, and disrupting it should impair awareness or metacognition without impairing first-order discrimination.

    The SEP cites Lau and Passingham (2006): subjective reports dropped while discrimination stayed matched, linked to dorsolateral PFC. It also cites Rounis et al. (2010): TMS to PFC affected meta-visual awareness without harming task performance. Against this, a later theta-burst TMS attempt (Bor et al. 2017) did not replicate the Rounis effect. The SEP also reports that extensive PFC damage, including lobotomy, does not abolish experience, and that this lesion argument has met substantial pushback. plato.stanford.edu

  • Mixed

    Awareness and first-order processing can come apart, so discrimination can survive without awareness (blindsight). The theory explains this as a first-order representation lacking its higher-order counterpart.

    Lau's PRM explains blindsight as an intact first-order representation with no higher-order reliability tag (Closer to Truth summary). The pattern itself is well established, but first-order theorists interpret it differently, for example as degraded first-order signals. So the data do not by themselves favour higher-order theory. loc.closertotruth.com

  • Untested

    HOROR and PRM make predictions about the brain basis of visual consciousness that differ from Recurrent Processing Theory, tested in a Templeton adversarial collaboration.

    TWCF project 0567 (director Biyu Jade He, with Chalmers and Block; December 2020 to December 2023) designed two experiments so each side could falsify the other. The project page reports no results from those experiments. templetonworldcharity.org

  • Untested

    Rich and sparse variants, and relational and non-relational variants, make different predictions. Tests include psychophysical adaptation, imaging of perception versus imagination, and hypnotic suggestion as a test of misrepresentation.

    ETHOS launched in summer 2024, with designs due by July 2025: Denison and Peters on adaptation, Dijkstra with Morales on perception versus imagination, Dienes on hypnosis. The March 2025 post reported no results. templetonworldcharity.org

  • Mixed

    Prefrontal involvement in consciousness, a commitment HOT shares in part with global workspace theory, is tested indirectly by the Cogitate IIT-versus-GNWT collaboration.

    Cogitate (Nature, 2025) found information about conscious content in inferior frontal as well as posterior cortex. It also found limited representation of some conscious dimensions in prefrontal cortex, and the authors report no ignition at stimulus offset. Cogitate tested IIT and GNWT, not HOT, so any implication for HOT is indirect. comdig.cssociety.org

Strongest objections

  • Ned Block (2011, as reported in the SEP). There is no reason to expect a higher-order thought that one is in pain to have pain's negative valence and high arousal. Since those are central to conscious pain, a targetless higher-order thought would not be enough for it. plato.stanford.edu
  • Scott Sturgeon (misrepresentation objection, as reported in the SEP). A malfunctioning higher-order monitor could pair a red percept with a 'seems orange' higher-order state. The subject would then judge 'it's red' and 'it seems orange' together, which clashes with the immediacy of our awareness of our own experiences. Related targetless and misrepresentation problems are raised for actualist HOT. plato.stanford.edu
  • Over-intellectualization critics (animals and infants). Actualist and dispositionalist HOT seem to require mental-state concepts that many animals and infants may lack, which would wrongly deny them consciousness. plato.stanford.edu
  • Prefrontal localists (SEP survey). Lesion evidence suggests PFC is not needed for basic sensory consciousness: extensive PFC damage does not abolish experience or affect colour and shape experience. That counts against prefrontal versions of HOT. The SEP notes this argument has faced substantial pushback. plato.stanford.edu
  • Ned Block, overflow debate. Block argues that phenomenal consciousness overflows cognitive access, which would make access-linked higher-order states the wrong marker of consciousness. Brown disputes the overflow reading. philosophyofbrains.com
  • Ginger Campbell, Brain Science Podcast host, on Lau's book. She found the PRM argument unpersuasive. Her main objections: it does not explain the qualitative character of experience, and it limits consciousness to a small set of mammals. This is a popular-venue opinion, weaker than the scholarly critiques above. brainsciencepodcast.com

How it has been changing

  1. 2026-01-09Substantive revision of the Stanford Encyclopedia entry on higher-order theories. It discusses the prefrontal debate (Lau and Passingham, Rounis et al., lesion evidence) and adversarial collaborations. plato.stanford.edu
  2. 2025-03-31A Templeton post says ETHOS designs are due by July 2025, with experiments by Denison and Peters (rich vs sparse), Dijkstra and Morales (perception vs imagination) and Dienes (hypnosis and misrepresentation). No results yet. templetonworldcharity.org
  3. 2025The Cogitate Nature paper challenges key tenets of IIT and GNWT, including limited prefrontal representation of some conscious dimensions. Higher-order theories were not directly tested. comdig.cssociety.org
  4. 2024-2025Rosenthal's 'Thought, consciousness, and the given' (European Journal of Philosophy, doi 10.1111/ejop.13039) connects Sellars on thought with his HOT account. davidrosenthal.org
  5. 2024ETHOS launches in summer 2024, led by Fleming and Cleeremans. It tests variants within the higher-order family against each other instead of treating the family as one view. templetonworldcharity.org
  6. 2023-12TWCF project 0567 (HOROR and PRM versus Recurrent Processing Theory) reaches its listed end date, December 22, 2023. The project page reports no experimental outcomes. templetonworldcharity.org
  7. 2023-08Butlin, Long, Chalmers and colleagues include higher-order theory indicators, among those from several theories, in a rubric for assessing AI consciousness. Conclusion: no current system is a strong candidate, but there is no obvious technical barrier. arxiv.org

Open questions inside the theory

  • What is the reality-monitoring signal: a dedicated discriminator (Lau's PRM) or a read-out of signal strength in a higher-order state space (Fleming's HOSS)?
  • Can higher-order states strongly misrepresent first-order states, as non-relational theories hold, and can hypnosis show it?
  • Are higher-order contents rich or sparse?
  • Will ETHOS results separate the variants, or is the family too flexible to falsify?
  • Is prefrontal involvement needed for basic sensory consciousness, or only for report and confidence?
  • Do animals and infants have the required higher-order representations, and could current or near-term AI systems meet the higher-order indicators?
IllusionismBrains misrepresent their own workings so that they seem to have ineffable inner qualities. Explain why the illusion arises (the illusion problem) and the hard problem dissolves.

Illusionism. Keith Frankish, Daniel Dennett, Nicholas Humphrey, Michael Graziano, Francois Kammerer, Susan Blackmore, Georges Rey, Maxim Gorbachev. Last checked 15 Nov 2026.

Core claims

  • Phenomenal consciousness, conceived as intrinsic, ineffable, private qualities, is an illusion. Illusionists do not deny that we have experiences in a functional sense: we perceive, feel and introspect. What they deny is that experiences have phenomenal properties. The task is to explain why experiences seem to have them ('quasi-phenomenal' representations). Frankish's 2016 Journal of Consciousness Studies target article, with commentaries, set out the programme. susanblackmore.uk
  • Illusionism replaces the hard problem with the 'illusion problem': explaining why and how introspection leads us to judge that we have phenomenal properties. Chalmers's meta-problem frames the same question, and he notes that illusionists expect solving it to solve or dissolve the hard problem. philarchive.org
  • In his 2016 JCS commentary, Dennett called illusionism the obvious default theory of consciousness: the mundane possibility of illusion should be explored before exotic positions that offer few empirical predictions. Frankish replies to this and the other commentaries. Dennett's earlier work includes Quining Qualia (1988) and heterophenomenology, which treats subjects' reports as data to interpret, not as infallible access to inner facts. keithfrankish.github.io
  • Dennett's 2017 From Bacteria to Bach and Back describes consciousness as a user-illusion, an interface that lets us monitor and communicate our states without access to the underlying machinery. It sits inside an evolutionary account of competence without comprehension. nature.com
  • Graziano's attention schema theory is an openly illusionist neuroscience theory. His 2024 eNeuro paper argues illusionism is an umbrella framework and that an illusionist layer can be added to theories such as IIT and global workspace theory. eneuro.org
  • Frankish separates strong illusionism (phenomenal consciousness does not exist; it only seems to) from weak illusionism (it exists but lacks some properties we attribute to it, such as ineffability). He holds that only the strong version is illusionism proper, since the weak one is a form of realism. The 'distortion thesis' says introspection misrepresents neural processes as having phenomenal properties. oro.open.ac.uk

How it explains the evolution of consciousness

Illusionists give the illusion a biological origin but differ on its function. Dennett and Graziano treat the introspective self-model as adaptive. For Graziano, an attention schema helps control attention and model other minds. For Dennett, a user-illusion lets organisms monitor and report their own states. Humphrey argues that phenomenal seeming evolved because it makes life feel rich and worth living. Some illusionists leave open whether the illusory character is itself selected or a by-product of introspective representation. Dennett adds a cultural layer, with memes and language shaping human comprehension. The theory gives consciousness no cosmic role: it is a late biological and cultural product, and the interesting question is why brains came to represent and report it as they do.

Predictions and where they stand

  • Supported

    Without awareness of a stimulus, attention to it should still be possible but less stable and less controlled (attention schema theory).

    Wilterson et al. (2020) report that when subjects were unaware of a cue, it still captured attention, but endogenous control of attention was impaired. This fits AST's predictions. The results come from the theory's own lab, so independent replication is limited. grazianolab.princeton.edu

  • Not yet testable

    A solution to the meta-problem (why we judge consciousness to be hard to explain) will be a purely functional account that also dissolves the hard problem.

    Chalmers calls the meta-problem tractable but argues solving it does not by itself settle the hard problem. Kammerer, himself an illusionist, argues current illusionist theories cannot yet explain why illusionism seems so absurd. No agreed full solution exists. philarchive.org

  • Untested

    Introspection systematically distorts neural processes into apparent phenomenal properties, so specific patterns of distortion can be mapped and tested.

    Gorbachev and Frankish (2025) classify eight possible distortion patterns by positiveness, conceptuality and dispellability. They say empirical work is needed and do not defend one pattern. oro.open.ac.uk

Strongest objections

  • Galen Strawson. Calls the denial of consciousness the silliest claim ever made. Argues that to seem to have experience is already to have experience, so the illusion cannot be what illusionists say. Dennett replied in the same venue. nybooks.com
  • Francois Kammerer (internal critique by an illusionist). Poses the illusion meta-problem: illusionists must explain why their own thesis seems so counterintuitive, and he argues no current illusionist theory does. research.dial.uclouvain.be
  • Jesse Prinz. In the 2016 JCS commentaries, argued that illusionism either collapses into realism or introduces a puzzle similar to the hard problem. susanblackmore.uk
  • David Chalmers. Not an illusionist, though he has said that if he were a materialist he would be one. Published 'Debunking arguments for illusionism about consciousness' (2020) and argues the meta-problem is a problem for everyone. stafforini.com

How it has been changing

  1. 2025-04-30The Cogitate adversarial collaboration (Nature, 256 participants) substantially challenged key tenets of IIT and GNWT. Illusionism makes no distinctive neural prediction there: Graziano holds it can be layered onto either theory, so the result neither confirms nor refutes it. bernstein-network.de
  2. 2025Gorbachev and Frankish, 'Illusionism and the distortion thesis' (Synthese), maps eight possible patterns of introspective distortion and calls for empirical work. The programme is moving from defending the view to specifying it. oro.open.ac.uk
  3. 2025Frankish's chapter 'Consciousness is not what it seems' (Wiley-Blackwell) presents illusionism as a post-Cartesian paradigm. keithfrankish.github.io
  4. 2024-10Graziano's 'Illusionism Big and Small' (eNeuro) recasts illusionism as an umbrella framework covering several hypotheses, including illusionist versions of IIT and global workspace theory. Elsewhere he prefers 'caricature' to 'illusion'. eneuro.org
  5. 2024-04Daniel Dennett died on 19 April 2024 at 82. The main philosophical defence of illusionism now rests with Frankish, Graziano, Humphrey and a younger generation. ncse.ngo
  6. 2024Frankish published 'The Ethical Implications of Illusionism' in Neuroethics, arguing illusionism does not undercut the ethical value of consciousness and may widen moral concern. Kammerer calls this the normative challenge and has proposed an ethics that does not rest on sentience. eprints.whiterose.ac.uk

Open questions inside the theory

  • What exactly is the introspective distortion, and can it be tested empirically?
  • Can illusionism explain why it seems so counterintuitive (Kammerer's illusion meta-problem) without circularity?
  • Was the illusory character of introspection selected for a function, or is it a by-product?
  • Does the 'seeming' that illusionists rely on smuggle phenomenality back in (Strawson)?
  • What follows for ethics and for judging AI systems, if moral status cannot rest on phenomenal consciousness?
  • Can illusionism yield distinctive predictions beyond the rival theories it can be bolted onto?

Prediction and staying alive

Experience grows out of a living body predicting and regulating itself.

Free energy and affectLiving systems persist by minimising prediction error about their own states. Solms argues that the felt side of that regulation, affect generated in the upper brainstem, is what consciousness fundamentally is.

Free energy principle, active inference and affect-first consciousness. Karl Friston, Mark Solms, Anil K. Seth, Jakob Hohwy, Maxwell Ramstead, Christopher Whyte, Ryan Smith, Jonathan Shock. Last checked 15 Nov 2026.

Core claims

  • Solms and Friston (2018, Journal of Consciousness Studies): the function of consciousness is free energy minimisation, realised in dual aspects. Subjectively it is felt as affect; objectively it appears as arousal that modulates postsynaptic gain. They argue that the two insights together specify sufficient conditions for attributing feeling to a system. discovery-pp.ucl.ac.uk
  • Solms (2019, Frontiers in Psychology) names affect as the elemental form of consciousness. Deviation from a homeostatic settling point (increasing uncertainty) is felt as unpleasure, and a return toward it as pleasure. Cognition such as perception and memory can run unconsciously; what consciousness adds is felt value. pmc.ncbi.nlm.nih.gov
  • Solms concludes that consciousness is generated in the upper brainstem (the reticular activating system, the parabrachial nuclei and the periaqueductal gray), not the cortex, and treats the cortex as stabilising conscious contents. He cites decorticate animals and children with hydranencephaly as evidence that consciousness persists without cortex. pmc.ncbi.nlm.nih.gov
  • Solms (2019) also proposes, as theory rather than an experimental finding, that conscious reconsolidation is 'predictive-work-in-progress': prediction errors return long-term memory traces to working memory, where they can be updated. pmc.ncbi.nlm.nih.gov
  • Ramstead, Albarracin, Kiefer, Klein, Fields, Friston and Safron (v1 April 2023, v4 January 2024) derive an 'inner screen' model of consciousness directly from the FEP. It draws on the quantum information-theoretic version of the FEP and the brain's nested, hierarchical neuroanatomy. arxiv.org
  • Whyte, Corcoran, Robinson, Smith, Moran, Parr, Friston, Seth and Hohwy (2024) argue that existing active-inference models of consciousness share a small set of theoretical commitments, and that these amount to a 'minimal (and testable)' theory of consciousness. arxiv.org

How it explains the evolution of consciousness

The theory gives no standalone evolutionary story, but its logic is biological. Any system that resists dissipation must model and regulate its own states. Solms reads consciousness as the felt layer of that regulation, anchored in an old brainstem system, so on his view it should be evolutionarily ancient and widely shared among vertebrates rather than a late cortical add-on. He cites Merker (2007), who argued that the upper brainstem kept a key role in consciousness across mammals. Friston's physics-level claim is more general: any particular system with a Markov blanket can be described as if it infers. Critics say this makes it too broad to say when consciousness appears, and Veit and Browning argue the framework has not yet tied free energy minimisation to evolutionary fitness. Solms's engineering programme tries to pin down the conditions by building agents with self-set homeostatic needs.

Predictions and where they stand

  • Mixed

    Consciousness survives loss of the cortex but not loss of the upper brainstem: decorticate animals and children with hydranencephaly remain conscious, while small brainstem lesions abolish consciousness.

    Merker (2007) presented behavioural evidence for consciousness without cortex, and the commentary it drew was mixed. Seth, reviewing The Hidden Spring, finds the evidence from children born with little or no cortex unconvincing, since assuming they lack consciousness is no safer than assuming they have it. He also notes that research on recovery from coma and vegetative states points to cortical networks. cambridge.org

  • Mixed

    An artificial agent with self-set homeostatic needs and felt uncertainty about meeting them would show behaviours taken as evidence of feeling in animals, such as hedonic place preference.

    Solms, Grimbly, Shock, Rosman and colleagues (arXiv 2609.03883, 3 September 2026) report that a simple, fully deterministic agent shows hedonic place preference. This shows the behaviour can be produced. Whether it indicates feeling is the disputed step. It is a preprint, not peer-reviewed. deeplearn.org

  • Supported

    Active-inference agents with multiple bodily needs survive better when sensory (likelihood) precision is routed toward the most threatened need, especially when combined with reweighting of preferences.

    In 120-step gridworld episodes, survival was 0.56 for the uniform baseline, 0.67 with preference reweighting, 0.83 with precision routing and 0.94 with both. The tests covered 83 layouts: 11 hand-designed (8 seeds each) plus 72 pre-registered procedural layouts. This supports an engineering claim about interoceptive prioritisation, not a claim that the agents are conscious. Source: arXiv 2610.18617 (15 October 2026), a single preprint, read via a secondary summary. theconsciousness.ai

  • Not yet testable

    Active-inference models of consciousness should converge on a minimal shared theory that can be tested against rival theories.

    The 2024 paper calls its theory 'minimal (and testable)', but its abstract gives no specific experimental designs. No head-to-head adversarial test of it was found. arxiv.org

  • Challenged

    The FEP should yield a nontrivial physics of living systems that can be checked against real systems.

    Aguilera et al. (2022) find that even in simple linear stochastic systems, the FEP's requirements are restrictive and unusual for living systems. Critics argue that the FEP's generality makes it effectively unfalsifiable. Andrews replies that this is a category error, because the FEP is a modelling framework rather than an empirical hypothesis. pmc.ncbi.nlm.nih.gov

Strongest objections

  • Anil Seth. In a Times Higher Education review of The Hidden Spring, Seth praises the central argument that feelings and bodily regulation are core to consciousness. He finds the brainstem evidence from children with little or no cortex unconvincing and the free energy exposition hard to follow, and he is unconvinced by the treatment of the hard problem and by the blueprint for a conscious machine. timeshighereducation.com
  • Lionel Naccache. In a Journal of Consciousness Studies commentary on The Hidden Spring, he discusses where his views converge with and diverge from Solms's, from the standpoint of global neuronal workspace theory. ouci.dntb.gov.ua
  • Charles Fernyhough. His review says Solms is oddly silent about what consciousness is like and dismisses competing theories quickly. charlesfernyhough.com
  • Biehl, Pollock and Kanai. In a technical critique, they argue that crucial steps in the free energy argument about Markov blankets need additional, unstated assumptions, and that published definitions of the blanket are not equivalent. pmc.ncbi.nlm.nih.gov
  • Aguilera, Millidge, Tschantz and Buckley. They tested the FEP in the simplest systems and found its conditions restrictive and unusual for living systems, which limits how informative it is. pmc.ncbi.nlm.nih.gov
  • Veit and Browning. In 'Life, mind, agency: why Markov blankets fail the test of evolution' (Behavioral and Brain Sciences, 2022), they argue that FEP proponents overlook evolutionary dynamics and give no convincing adaptive account linking free energy minimisation to fitness. They see this as a central obstacle to the claim of unifying life and mind. researchonline.lse.ac.uk
  • Mel Andrews (a defence that answers the unfalsifiability charge). Andrews argues that demands for the FEP to be falsifiable rest on a category error, since it is a modelling framework. Critics counter that this leaves the principle without empirical content. philarchive.org

How it has been changing

  1. 2026-10Grimbly, Shock, Rosman, Smith and Solms post arXiv 2610.18617, a simulation study of two routes to interoceptive prioritisation (preference reweighting and likelihood precision) in active-inference agents with multiple needs. Read via a secondary summary. theconsciousness.ai
  2. 2026-09Solms and the Cape Town group post arXiv 2609.03883: a simple, deterministic agent with a felt-uncertainty architecture shows hedonic place preference. They discuss what this implies for the physical basis of consciousness and for free will. deeplearn.org
  3. 2025-04 (online); 2026 (BBS vol. 49 with commentaries)Seth's target article in Behavioral and Brain Sciences argues for biological naturalism: consciousness is tied to living systems, so AI consciousness is unlikely on current trajectories. About 50 commentaries followed, some disputing whether predictive processing or the FEP is what matters. Seth is a prominent voice close to the FEP programme. resolve.cambridge.org
  4. 2025-04The Cogitate Consortium's adversarial collaboration, published in Nature, challenges core tenets of both integrated information theory and global workspace theory. It did not test FEP or affect-first views. idw-online.de
  5. 2024-11/12Solms and Friston give talks on engineering an artificial consciousness, and Solms describes his 'felt uncertainty' theory and early work on an engineered agent. The only evidence is event listings. npsa-association.org
  6. 2024-10 (v1), 2025-06 (v2)Whyte, Friston, Seth, Hohwy and five co-authors propose a minimal theory of consciousness implicit in active inference, an attempt to make the framework comparable with other theories. arxiv.org
  7. 2023-04 to 2024-01Ramstead, Friston, Safron and colleagues post and revise the 'inner screen' model (v1 15 April 2023, v4 2 January 2024). It is a consciousness model derived from the FEP, separate from Solms's affect route. arxiv.org

Open questions inside the theory

  • Does felt uncertainty in a simulated agent amount to feeling, or only a functional analogue? What test would separate the two?
  • Given its generality, can the FEP make any claim that distinguishes conscious from non-conscious self-organising systems?
  • Is the upper brainstem sufficient for affect, and how much do hydranencephaly cases with cortical remnants weaken the evidence?
  • How would an affect-first account fare in an adversarial collaboration against workspace, higher-order or recurrent-processing theories?
  • Do the inner screen model and Solms's brainstem model make compatible or competing predictions?
Predictive processingThe brain constantly predicts its inputs, and experience is the "controlled hallucination" those predictions produce; Seth argues the deepest predictions serve keeping the body alive.

Predictive processing / the beast-machine view. Anil Seth, Andy Clark, Jakob Hohwy, Karl Friston, Manos Tsakiris. Last checked 15 Nov 2026.

Core claims

  • Perception is top-down inference: experience of the world is a 'controlled hallucination' kept in check by sensory prediction error (Seth's popular framing). cbc.ca
  • Predictive processing is not itself a theory of consciousness, but it can give a systematic basis for finding neural correlates of consciousness (Hohwy and Seth, 2020, who concede it is not in itself a theory and hope parts may coalesce into one). philosophymindscience.org
  • Selfhood is grounded in 'instrumental interoceptive inference': the felt self arises from predictions about the body's internal state used to regulate it, so experience arises because of, not in spite of, our nature as 'beast machines' (Seth and Tsakiris, Trends in Cognitive Sciences, 2018). sro.sussex.ac.uk
  • Consciousness has more to do with being alive than with being intelligent; Being You (2021) ties it to physiological regulation. Seth says this is less a rival theory than an approach to the 'real problem' of consciousness. mind-foundation.org
  • Computation alone is probably not sufficient for consciousness; real artificial consciousness is unlikely on current trajectories and more plausible if AI becomes more life-like (biological naturalism, Behavioral and Brain Sciences, 2025). sussex.figshare.com
  • Clark's wider framework: minds are prediction machines whose predictions also drive action, reshaping the world they then sense (Surfing Uncertainty, 2016); The Experience Machine (2023) applies it to pain and mental health. philosophyofbrains.com

How it explains the evolution of consciousness

In Seth's account, nervous systems evolved to regulate the body, so the most basic experiences are of being a living organism (mood, emotion, bodily self); perception of the outside world is an elaboration of the same predictive machinery. Consciousness is therefore tied to life and metabolism, not to intelligence or computation. There is no cosmic account: it is a biological, organism-level view, and Seth argues against substrate-independent computationalism.

Predictions and where they stand

  • Mixed

    Interoceptive inference: predictions about bodily signals shape emotion and the felt self, so heartbeat-related brain signals should show precision-weighted prediction effects.

    Seth and Friston's 2016 review said direct empirical evidence for interoceptive inference was still to be found. Banellis and Cruse (Cerebral Cortex Communications, 2020) found heartbeat-evoked potentials reflecting cardiac-driven predictions of sounds, consistent with attentional precision; a 2021 follow-up preprint from the same group replicated the prediction effect but reported results not consistent with precision-weighting. biorxiv.org

  • Mixed

    Sensory responses to expected stimuli should be suppressed, as if prediction error is being minimised (repetition and expectation suppression).

    Predictive accounts explain repetition suppression via prediction error, but reviews note attention, recognition and learning can inflate, abolish or reverse the effect, and non-predictive models explain subsets of the data. pmc.ncbi.nlm.nih.gov

  • Challenged

    Predictive processing can discriminate among and constrain theories of consciousness (the NCC programme).

    Schlicht and Dołęga argue it is too vague to individuate neural mechanisms at a fine enough scale to yield breakthroughs in the search for neural correlates of consciousness. philosophymindscience.org

  • Untested

    Large individual differences in how expectations shape perception and imagery (Perception Census, launched 2022).

    Citizen-science study led by Seth and Macpherson seeking tens of thousands of participants; published results were not located in the sources checked. sussex.ac.uk

  • Not yet testable

    Conscious AI built on computation alone is unlikely; life-like substrates matter.

    A philosophical and theoretical claim; commentators (e.g. Kleiner) ask that biological naturalism be defined more explicitly because many properties discussed may not be unique to living systems. sussex.figshare.com

Strongest objections

  • Schlicht and Dołęga. Predictive processing is too vague to individuate neural mechanisms at a fine enough scale, so it is unlikely to yield breakthroughs in the search for neural correlates of consciousness. philosophymindscience.org
  • Dołęga and Dewhurst. Problems of unconscious perception and of the account of consciousness threaten to trivialise or sharply limit Hohwy's prediction-error-minimisation account of consciousness. link.springer.com
  • Doerig, Schurger and Herzog. Of 13 theories assessed against hard empirical criteria, none emerged unscathed; predictive processing, like global workspace and higher-order theories, is vulnerable to the small and large network argument. infoscience.epfl.ch
  • Biehl, Pollock and Kanai. Technical critique of the free energy principle that underlies much of the framework: the Free Energy Lemma, taken at face value, is wrong in general, and Markov blanket definitions are not equivalent (they note this does not rule out the broader ideas). arxiv.org
  • Veit and Browning. Free energy and Markov blanket accounts neglect evolutionary dynamics: they offer no convincing adaptive story linking free-energy minimisation to fitness and fail to pick out the organismal boundary that matters for evolution (Behavioral and Brain Sciences commentary, 2022; relevant to Seth's life-first story). researchonline.lse.ac.uk

How it has been changing

  1. 2026-01Seth's essay 'The Mythology of Conscious AI' wins the English-language category of the 2025 Berggruen Prize Essay Competition (about 3,000 submissions) and is published in Noema, arguing consciousness is more likely a property of life than of computation. berggruen.org
  2. 2025-04-30Cogitate adversarial collaboration (Nature, n=256) substantially challenged key tenets of GNWT and IIT. Predictive processing was not a tested party, so it gains no direct support, but it shows the testing culture the field is moving toward. research.birmingham.ac.uk
  3. 2025Seth's 'Conscious artificial intelligence and biological naturalism' appears as a Behavioral and Brain Sciences target article with commentaries, moving the beast-machine view toward an explicit argument that consciousness may require life. resolve.cambridge.org
  4. 2025Clark's 'Extending Minds with Generative AI' (Nature Communications 16:4627) applies his extended-mind side of predictive processing to AI tools, a contrast with Seth's life-first view of consciousness. ideas.repec.org
  5. 2023Clark's The Experience Machine popularises the framework and applies it to pain and mental health. books.apple.com
  6. 2022-05Seth and Bayne's Nature Reviews Neuroscience review places predictive processing alongside higher-order, global workspace and IIT theories and calls for iterative testing and comparison. sro.sussex.ac.uk
  7. 2022Perception Census launched by Seth and Macpherson to study perceptual diversity and how expectations shape what people see. sussex.ac.uk

Open questions inside the theory

  • Is predictive processing a theory of consciousness, or only a framework that other theories can use? Hohwy and Seth concede the latter for now.
  • What would a decisive, preregistered adversarial test of predictive processing against GNWT, IIT or higher-order theories look like?
  • Is life (metabolism, allostasis) necessary for consciousness, or only the control-oriented prediction that life usually implements?
  • Can interoceptive inference be tested directly, given mixed heartbeat-evoked potential results?
  • Does the free energy principle, which supplies much of the formal backbone, survive technical and evolutionary critiques?
  • How do cortical predictive accounts fit with views that locate core affect in the brainstem (e.g. Solms)?

Physics underneath

Mind depends on, or is described by, quantum physics and observers.

Orch ORConsciousness is proposed to arise from quantum computations in brain microtubules that end in gravity-linked collapse events, each one a moment of experience.

Orchestrated Objective Reduction (Orch OR). Roger Penrose, Stuart Hameroff, Michael Wiest, Jack Tuszynski. Last checked 15 Nov 2026.

Core claims

  • Hameroff and Penrose's 2014 review (Physics of Life Reviews) states the theory: consciousness arises from quantum computations in microtubules that end through Diosi-Penrose objective reduction, with the events orchestrated by neural inputs and tied to moments of awareness or choice. experts.arizona.edu
  • Penrose's Godel-based argument (The Emperor's New Mind, 1989; Shadows of the Mind, 1994) holds that human mathematical understanding is non-computable, which motivates a physics beyond standard quantum mechanics. Orch OR joins this argument to Hameroff's microtubule hypothesis. en.wikipedia.org
  • Hagan, Hameroff and Tuszynski (2002) argued that Tegmark modeled the wrong mechanism, and that corrected decoherence times are around 10^-5 to 10^-4 s, extendable by shielding mechanisms such as ordered water and counterion layers. arxiv.org
  • Kalra et al. (2023, with Hameroff, Tuszynski and Penrose among coauthors) measured tryptophan fluorescence and report that electronic excitation migrates about 6.6 nm along microtubules, further than conventional Forster theory accounts for. Etomidate and isoflurane reduced the migration. pmc.ncbi.nlm.nih.gov
  • Anesthetic action on microtubules is read as a causal link to unconsciousness. Wiest's 2025 paper argues that a microtubule substrate is the best fit to the anesthetic data, that classical synaptic accounts fit worse, and that this substrate is specifically predicted by Orch OR. pmc.ncbi.nlm.nih.gov

How it explains the evolution of consciousness

Penrose's framing makes consciousness a feature of fundamental physics, not a late biological add-on. Objective reduction events are proposed to happen throughout the universe as 'proto-conscious' moments, and biology orchestrates them into full consciousness. Hameroff (1998, 'Did Consciousness Cause the Cambrian Evolutionary Explosion?', in Toward a Science of Consciousness II, MIT Press) proposed that Orch OR consciousness first appeared in simple animals around the Cambrian explosion, about 540 million years ago, and helped drive it. More recently Hameroff has argued a speculative 'consciousness came before life' position, in which proto-conscious quantum events helped make life possible.

Predictions and where they stand

  • Challenged

    Quantum coherence can survive long enough in brain microtubules for the proposed computations, against Tegmark's estimate of about 10^-13 s.

    Tegmark (2000) calculated decoherence far shorter than relevant brain timescales. Hagan et al. (2002) revised to 10^-5 to 10^-4 s with extra shielding mechanisms. Those mechanisms have not been shown in living cells. Koch and Hepp (2006) judged quantum computation in the warm wet brain implausible. Reimers et al. (2014) argued the proposal lacks a justified account of long-lived quantum states in microtubules. arxiv.org

  • Challenged

    Parameter-free Diosi-Penrose gravity-related collapse should produce spontaneous radiation detectable above background.

    Donadi et al. (Nature Physics, online 2020, vol. 17, 2021), using a germanium detector at Gran Sasso, ruled out the natural parameter-free version of the Diosi-Penrose model. This constrains that form of the collapse model, not gravity-related collapse in general, and not Orch OR's biological claims directly. nature.com

  • Supported

    Microtubule-active drugs should alter anesthetic-induced unconsciousness.

    Khan et al. (2024, Wiest lab) found rats given the microtubule stabilizer epothilone B took about 69 seconds longer to lose righting reflex under isoflurane, as predicted. Support is preliminary: one drug, one anesthetic, one species, from a lab sympathetic to the theory, no independent replication found, and a classical microtubule role is not excluded by the result. eneuro.org

  • Mixed

    Microtubules should support unusually far-reaching, collective excitation transfer and quantum optical effects (such as superradiance) relevant to coherence.

    Kalra et al. (2023) report long-range energy migration along microtubules that Forster theory does not fully explain. Babcock et al. (2024) report theoretical UV superradiance in tryptophan networks with a supporting fluorescence-yield experiment. Both are in vitro or theoretical, and neither shows quantum coherence at the timescales or in the living-brain setting Orch OR needs. pubs.acs.org

  • Not yet testable

    Quantum entanglement across microtubules in different neurons underlies binding of experience.

    A 2025 Frontiers in Human Neuroscience opinion piece (Sergi et al.; edited by Wiest, reviewed by Hameroff) proposes a quantum-classical framework and itself states that entanglement of microtubules among neurons 'has yet to be proven'. frontiersin.org

Strongest objections

  • Max Tegmark. Calculated neural decoherence times of about 10^-13 to 10^-20 s, vastly shorter than the 0.001 to 0.1 s of neural and microtubule dynamics, so brain cognitive degrees of freedom behave classically. arxiv.org
  • Christof Koch and Klaus Hepp. In Nature (2006) they argued the 'wet and warm' brain makes quantum computation implausible and there is little reason to invoke quantum mechanics for higher brain function. pubmed.ncbi.nlm.nih.gov
  • Reimers and colleagues. Their 2014 commentary on the Hameroff-Penrose review argues the revised proposal is not scientifically justified, lacking a precise account of the qubit states, how they entangle, and how coherence is held over the needed time. Hameroff and Penrose replied in the same journal that the critics' microtubule model was oversimplified. newdualism.org
  • Solomon Feferman, Marvin Minsky, John Searle and others. The Godel-based argument motivating the theory is widely held to fail: it requires that humans can know their own mathematical reasoning is sound and consistent, which critics deny, and human reasoning may be inconsistent or unknowably algorithmic. Feferman reviewed Shadows of the Mind critically on these lines. calculemus.org

How it has been changing

  1. 2025-09Sergi et al. in Frontiers in Human Neuroscience propose a quantum-classical mathematical (Hamiltonian plus complexity-theory) framework to make Orch OR more predictive and testable, and acknowledge cross-neuron microtubule entanglement is unproven. frontiersin.org
  2. 2025-05Wiest publishes 'A quantum microtubule substrate of consciousness is experimentally supported and solves the binding and epiphenomenalism problems' in Neuroscience of Consciousness, a synthesis arguing the evidence is predicted by Orch OR. The experimental support is the earlier anesthetic and microtubule work. pmc.ncbi.nlm.nih.gov
  3. 2024-08Khan et al. (Wiest lab) in eNeuro: epothilone B delays isoflurane-induced unconsciousness in rats by about 69 seconds, presented as support for microtubule-based theories. eneuro.org
  4. 2024-04Babcock et al. publish the tryptophan UV superradiance paper in J Phys Chem B, offered as a physical basis for collective quantum optical behavior in tubulin-like architectures. Neuroprotection and signaling readings are hypotheses. pubs.acs.org
  5. 2023-01Kalra et al. in ACS Central Science (online January, issue March 2023) report long-range electronic energy migration in microtubules, with Hameroff, Tuszynski and Penrose among coauthors; anesthetics reduced it. pmc.ncbi.nlm.nih.gov

Open questions inside the theory

  • Can any quantum coherence in microtubules be shown in living neurons at the millisecond timescales the theory needs, not just in vitro?
  • Does microtubule involvement in anesthesia require a quantum role, or does a classical cytoskeletal mechanism explain the epothilone result equally well?
  • Can a variant of Diosi-Penrose collapse survive the Gran Sasso constraint, and would it match what Orch OR needs?
  • Is there any independent, preregistered test of Orch OR comparable to the IIT versus GNWT adversarial collaboration?
  • How would cross-neuron entanglement be established, given the theory's use of it for binding?
Quantum observersAny physical system separable from its environment can be treated as a quantum observer reading a boundary. On this view the free energy principle is asymptotically equivalent to unitarity, and cognition is scale-free rather than special to brains.

Quantum-information observer theory (holographic screens, quantum FEP). Chris Fields, James F. Glazebrook, Antonino Marcianò, Michael Levin, Karl Friston. Last checked 15 Nov 2026.

Core claims

  • The free energy principle, restated in spacetime-background-free, scale-free quantum information theory, is asymptotically equivalent to the Principle of Unitarity. Generic quantum systems can be treated as observers, and as agents that assign meaning to their observations once the standard freedom-of-choice assumption is added. arxiv.org
  • The holographic principle follows from the quantum information theory of separable systems: interaction between a system and its environment is information exchange across a holographic screen. The authors relate this boundary to the Markov blanket of the FEP, and hold that measurement outcomes become comparable and meaningful through quantum reference frames. arxiv.org
  • 'Minimal physicalism' assumes nothing beyond quantum information theory, so it applies from molecules upward. The authors argue that integrated information, state broadcasting via small-world networks and hierarchical Bayesian inference emerge naturally in it. They also argue that the self-representation behind human autonoetic awareness can be traced back to the stress response in bacteria and other basal systems, and serves the same basic functions. pmc.ncbi.nlm.nih.gov
  • Neurons, and hierarchies of neurons, are modelled as active-inference systems built from hierarchies of quantum reference frames. arxiv.org
  • In systems that do active inference under the FEP, control flow can always be represented as tensor networks, and these could be implemented in topological quantum neural networks. arxiv.org
  • Self-representation by any finite physical system can at best be heuristic. A theorem extending Moore's and Rice's classical results to quantum systems shows that a system cannot use a quantum reference frame it implements to determine that frame's dimension or input/output function. Corollaries: components of a composite system cannot fully determine or deterministically control each other, and physical models are incomplete. pmc.ncbi.nlm.nih.gov

How it explains the evolution of consciousness

Observerhood is defined for any system with a boundary and a reference frame, so consciousness has no sharp origin. Cognition is scale-free, and the authors trace human self-representation back to the stress response in bacteria and other basal systems. They read evolution and development as free-energy-minimising reorganisation of reference frames across scales. Their 2022 paper 'The free energy principle induces neuromorphic development' (arXiv 2207.09734) argues that the FEP drives systems toward hierarchical, neuromorphic control structures. We found no cosmological account in their work of how observers first arose.

Predictions and where they stand

  • Untested

    Basal, non-neural systems such as bacteria have self-representation, through the stress response, that is continuous with and serves the same functions as the self-representation behind human autonoetic awareness.

    Offered as a conceptual derivation in the 2021 paper. No experiment designed to test it against rival theories was found. pmc.ncbi.nlm.nih.gov

  • Untested

    Cognition in any embodiment can be characterised as competency in navigating arbitrary spaces (transcriptional, physiological, morphological, behavioural). An experimenter can build an empirical model of a system's quantum reference frames by reverse-engineering its input-output behaviour, but can never fully determine those frames.

    Fields and Levin (2022), section 7.5, argue that full experimental determination of a system's reference frames is impossible and that the best available result is a provisional model in the experimenter's own frame. They point to methods borrowed from cognitive psychology but give no specific protocol tied to the framework's distinctive claims. pmc.ncbi.nlm.nih.gov

  • Not yet testable

    No finite system can fully represent itself, and the self-models a system holds cannot in general be validated by that system.

    This is a formal theorem within the authors' generic quantum-information model, not an empirical claim. No independent check of the proofs was found. pmc.ncbi.nlm.nih.gov

Strongest objections

  • Biehl, Pollock and Kanai. They identify technical errors in Friston's 2019 FEP derivation: the rewriting of the equations of motion for systems with Markov blankets is not equivalent, and different papers define the Markov blanket differently. They say this does not rule out the general ideas. The point bears on the quantum FEP indirectly, because it builds on FEP and Markov-blanket foundations. arxiv.org
  • Bruineberg, Dolega, Dewhurst and Baltieri. Markov blankets are used both as an epistemic tool for statistical inference ('Pearl blankets') and as a physical boundary between agent and environment ('Friston blankets'). The second use needs extra philosophical premises that the formalism alone cannot supply. This bears on identifying Markov blankets with physical holographic screens. researchers.mq.edu.au
  • Veit and Browning. Predictive expectations and fitness values do not obviously coincide, so the claim that free energy minimisation captures fitness is unjustified. Without an account of evolutionary dynamics, the FEP's ambition to unify the mind and life sciences falls short. eprints.soton.ac.uk
  • Btesh, Bramley and Lagnado. The loose definition of Friston blankets blurs the distinction between observational and interventional inference. Where a blanket falls depends on the modeller. bramleylab.ppls.ed.ac.uk

How it has been changing

  1. 2026-01Hartl, Pio-Lopez, Fields and Levin posted 'Remapping and navigation of an embedding space via error minimization' (20 Jan 2026). It extends the navigating-arbitrary-spaces framework into a substrate-independent principle that it applies to both biological collectives and AI systems such as transformers, diffusion models and neural cellular automata. arxiv.org
  2. 2025-09Fields, Glazebrook, Marcianò and Zappala posted 'Amplituhedra for generic quantum processes via the TQNN representation of UQC' (v1 24 Sep 2025, revised Feb and Jul 2026). It builds universal quantum computation inside topological quantum neural networks and proposes a correspondence with amplituhedra. The work is moving toward formal quantum-computation and topology results that are not directly about consciousness. arxiv.org
  3. 2024-02-24Fields, Glazebrook and Levin published 'Principled Limitations on Self-Representation for Generic Physical Systems' (Entropy 26(3), 194). The programme turns to formal limits: self-representation is at best heuristic, and components cannot fully determine or control one another. pmc.ncbi.nlm.nih.gov
  4. 2023-02'Control flow in active inference systems' was posted on 25 Feb 2023 by Fields, Fabrocini, Friston, Glazebrook, Hazan, Levin and Marcianò, and later published in two parts in IEEE Transactions on Molecular, Biological and Multi-Scale Communications (2023). It casts active-inference control flow as tensor networks implementable in topological quantum neural networks. arxiv.org

Open questions inside the theory

  • Can the quantum FEP and the holographic-screen formalism make any prediction that rival theories (IIT, global workspace, the classical FEP) do not also make?
  • Does treating any separable system with a boundary and reference frame as an observer explain consciousness, or only information processing?
  • Do the Pearl-versus-Friston blanket critiques undermine the identification of Markov blankets with physical holographic screens?
  • Given the authors' own result that a system's reference frames cannot be fully determined from outside, what empirical work could confirm or refute their reference-frame models of real biological systems?
  • How do the formal tensor-network, TQNN and amplituhedron results connect to claims about consciousness?

Selection at every scale

Agency and selection run from cells and molecules up to whole universes.

Assembly theoryObjects are defined by their construction history. A complex object found in many copies signals selection, which the theory treats as a measurable physical fact marking life and its derived forms.

Assembly theory. Lee Cronin, Sara Walker, Chris Kempes, Michael Lachmann, Cole Mathis, Stuart Marshall. Last checked 15 Nov 2026.

Core claims

  • The assembly index is the minimal number of recursive joining steps needed to build an object. Assembly combines that index with the object's copy number, as a measure of how much selection or causation was needed to produce the observed set of objects. pmc.ncbi.nlm.nih.gov
  • Selection and evolution can be quantified: high-index objects found in many copies are unlikely to occur without some information-driven selective process. The framework leaves physical law unchanged and changes what counts as an object. pmc.ncbi.nlm.nih.gov
  • The molecular assembly index can be measured experimentally, with tandem mass spectrometry, to flag molecules made by living systems. The authors calculated complexity for several million known molecules, and in their samples only living or biologically derived ones gave measured values above about 15. pmc.ncbi.nlm.nih.gov
  • The assembly index can also be estimated from infrared and NMR spectra as well as MS/MS, which makes it experimentally accessible without full structure determination. eprints.gla.ac.uk
  • Causation is a material property: assembly index plus copy number give a measurement framework for causation and contingency, and an assembly threshold that separates life and its agential, intelligent and technological derivatives from other structures. Novelty, contingency and open-endedness are treated as fundamental, with determinism emerging from selection along assembled lineages. arxiv.org
  • The assembly index is claimed to be formally distinct from Shannon entropy and compression-based measures such as Huffman coding and LZW, with computing it shown to be NP-complete. pmc.ncbi.nlm.nih.gov

How it explains the evolution of consciousness

Evolution is the accumulation of assembly along lineages. Selection is detected when high-assembly objects exist in abundance, and the environment acts as a kind of memory that keeps producing them. Cronin and Walker's 2026 preprint frames life, and its agential, intelligent and technological derivatives, as what lies above an assembly threshold, with determinism emerging from selection along lineages. Consciousness is not given its own account; at most it would fall under agency or intelligence as a derived form of life, and the core sources do not apply the theory to subjective experience.

Predictions and where they stand

  • Challenged

    Molecules with a molecular assembly index of about 15 or above, found in high copy number, are products of life or of a selective process.

    Hazen and colleagues calculated indices up to 21 (18 by the exact method) for abiotic mineral heteropolyanions such as ilmajokite and argued that 15 is not an unambiguous biosignature. Their indices are computed by hand from crystal structures, not measured. Walker, Mathis, Marshall and Cronin replied that experimentally measured indices are required to set the threshold and that comparing inorganic clusters with covalent molecules is a category error. The dispute is unresolved. pmc.ncbi.nlm.nih.gov

  • Supported

    The assembly index of an unknown molecule can be measured experimentally, by tandem mass spectrometry and also by IR and NMR, without full structure determination.

    Marshall et al. 2021 measured molecular assembly by MS/MS across a range of biological and abiotic samples. Jirasek et al. 2024 extended estimation to IR and NMR and compared the three techniques. This supports measurability. Independent replication of the biosignature threshold by other groups is not shown in these sources. pmc.ncbi.nlm.nih.gov

  • Mixed

    Assembly index is not reducible to Shannon entropy or compression, so it adds information beyond existing complexity measures.

    Kempes et al. 2025 give counterexamples against Huffman and LZW and an NP-completeness argument. Zenil and colleagues maintain it is equivalent to, or a weak form of, LZ-based algorithmic complexity. Both sides publish; no consensus. pmc.ncbi.nlm.nih.gov

  • Untested

    Assembly theory explains and quantifies selection and evolution, so it should detect selection in non-molecular systems as well as chemistry.

    The 2023 Nature paper is theoretical. Jaeger 2024 argues that the bias it detects can come from non-selective processes such as self-organization, kinetic barriers or drift, so detecting bias does not by itself show Darwinian selection. pmc.ncbi.nlm.nih.gov

Strongest objections

  • Uthamacumaran, Abrahao, Kiani, Zenil. Assembly pathway methods are a restricted form of Huffman-style encoding, fail on simple cases, and offer no advantage over existing complexity methods. arxiv.org
  • Abrahao, Hernandez-Orozco, Kiani, Tegner, Zenil. Assembly theory is a weak form of algorithmic complexity based on LZ compression and does not explain selection or evolution beyond known connections. arxiv.org
  • Johannes Jaeger. The approach has merit as a tool for detecting higher-level organization, but its metrics are special cases of classical complexity measures, its notion of 'selection' covers any constraint-driven bias, it cannot identify what caused the bias, and the title and abstract overstate the result. pmc.ncbi.nlm.nih.gov
  • Hazen, Burns, Cleaves, Downs, Krivovichev, Wong. Abiotic mineral heteropolyanions reach calculated assembly indices up to 21, so an index of 15 or more is not an unambiguous biosignature. pmc.ncbi.nlm.nih.gov

How it has been changing

  1. 2026-01Cronin and Walker post 'The Physics of Causation', framing causation as a material property measured by assembly and extending the threshold to agential, intelligent and technological derivatives of life; revised April 2026. arxiv.org
  2. 2025-09Kempes et al. in npj Complexity argue the assembly index is formally distinct from Shannon entropy and compression measures, and NP-complete to compute. pmc.ncbi.nlm.nih.gov
  3. 2024-06Walker, Mathis, Marshall and Cronin post a reply arguing experimentally measured assembly indices are required to set the threshold for life. arxiv.org
  4. 2024-04Jirasek et al. in ACS Central Science show assembly index estimation from IR, NMR and MS/MS. eprints.gla.ac.uk
  5. 2024-03Jaeger in the Journal of Molecular Evolution, and Zenil's group on arXiv, publish critiques arguing overstated novelty and equivalence to LZ compression. pmc.ncbi.nlm.nih.gov
  6. 2024-02Hazen et al. in J. R. Soc. Interface report abiotic heteropolyanions with calculated indices up to 21, challenging the index-15 threshold. pmc.ncbi.nlm.nih.gov
  7. 2023-10Sharma et al. publish 'Assembly theory explains and quantifies selection and evolution' in Nature, the main statement of the theory. pmc.ncbi.nlm.nih.gov

Open questions inside the theory

  • Is there a validated molecular assembly threshold that abiotic chemistry cannot cross, once inorganic and covalent systems are treated fairly?
  • Does the assembly index add anything beyond compression-based complexity measures in practice?
  • Can the framework detect selection outside chemistry, such as in technology, language or minds, and distinguish it from other sources of bias?
  • How would the theory say anything about subjective experience rather than agency and causation?
  • Can independent groups replicate the mass spectrometry biosignature results?
Cosmological selectionUniverses reproduce through black holes with slightly varied constants, so physics is tuned by selection. Gough adds that this favours universes that evolve intelligence able to make black holes.

Cosmological natural selection (Smolin) and its extension in The Egg and the Rock (Gough). Lee Smolin, Julian Gough, Leonard Susskind (critic), Alexander Vilenkin (critic), Ruediger Vaas (critic), Tony Rothman and George Ellis (critics). Last checked 15 Nov 2026.

Core claims

  • Smolin's hypothesis (first published in 1992, popularised in The Life of the Cosmos, 1997): black holes produce new universes with slightly altered constants. Universes whose constants favour more black holes leave more descendants, so such constants become common. scientificamerican.com
  • Smolin presents it as a falsifiable alternative to anthropic reasoning on a landscape of possible parameters, and sets out the conditions such a theory must meet to yield falsifiable predictions for doable experiments. arxiv.org
  • In Smolin's version life is incidental, 'just an incidental consequence of a universe having sufficient complexity to give rise to black holes' (Carr's 1997 review), not a target of selection. physicsworld.com
  • Gough's extension (the Blowtorch Theory): a wave of direct-collapse supermassive black holes forms first, before stars; their jets shock and enrich gas into stars and galaxies and carve out the cosmic web. Galaxies then make stellar black holes, and later heavy elements, planets, life and technology. theeggandtherock.com
  • Gough argues selection would favour universes that produce intelligence. Black holes are the most efficient energy source physics allows (he cites up to 42% mass-to-energy, against about 0.7% for fusion), so technological civilisations would make many small black holes, more than natural processes do. theeggandtherock.com
  • On consciousness, Gough says he deliberately leaves it out of The Egg and the Rock so the book can explain material reality with current physics. He speculates, and admits it is highly speculative, that entanglement dating from the Big Bang might give a material basis for a cosmic consciousness, and he plans a second book on it. theeggandtherock.com

How it explains the evolution of consciousness

Cosmic, not biological. Universes are the units of heredity and selection, and black holes are the reproductive organs. Mind arrives late. In Smolin's version it is only a side effect of the complexity needed for black holes. In Gough's it is a stage that raises reproductive output once intelligence makes small black holes. Gough explicitly defers the question of what consciousness is to a later, speculative book.

Predictions and where they stand

  • Challenged

    Neutron stars have a low maximum mass (about 1.5 solar masses under the Bethe-Brown kaon-condensation scenario Smolin relied on), since a lower limit means more black holes. Smolin wrote that any well-measured value above this 'would be troubling' and a pulsar above 2.5 solar masses 'would cleanly refute the theory'.

    In 1997 all well-measured neutron star masses were below 1.5 solar masses. Since then PSR J0740+6620 (about 2.08) and PSR J0952-0607 (2.35 ± 0.17, Romani et al. 2022) have been measured. Both are well above the low Bethe-Brown limit the prediction leaned on, but still below Smolin's 2.5 'clean refutation' threshold. arxiv.org

  • Challenged

    The constants of nature sit near a local maximum of black-hole production, so small changes to parameters should not increase the number of black holes.

    Vilenkin (2006) argued that raising the cosmological constant would increase black-hole production. Smolin replied that this rests on doubtful assumptions. Susskind (2004) also doubted that our laws maximise black-hole production. arxiv.org

  • Mixed

    JWST should find early galaxies dominated by, and formed around, supermassive black holes (Gough).

    Gough calls the Kocevski et al. spectral study of 341 'little red dots' 'a resounding triumph for three-stage cosmological natural selection.' The cosmologist Julian Munoz told him that photometry alone can't tell an ultramassive galaxy from a fainter one with a bright AGN, and that many little red dots look like AGN. That fits early black-hole-dominated objects but does not single out cosmological natural selection. The link to the theory is Gough's own interpretation. theeggandtherock.com

  • Untested

    About a trillion direct-collapse supermassive black holes form within the first 100 million years (probably the first 50 million), spinning near the Kerr limit, with jets carving voids and filaments.

    Stated by Gough in March 2025. He presents the Banados et al. 2024 Nature blazar (about 750 million years after the Big Bang), UHZ1 and Porphyrion as confirmations, but these are his own readings, not independent tests of these specific claims. theeggandtherock.com

  • Not yet testable

    A faint, overlapping 'popcorn' background of low-frequency gravitational waves from the brief direct-collapse era.

    Stated by Gough as a prediction. No detection or limit aimed at this signal was found. theeggandtherock.com

Strongest objections

  • Leonard Susskind (Edge exchange with Smolin, 2004). Three objections. Black holes do not lose information, so a baby universe could not inherit its parent's traits. Eternal inflation reproduces far more efficiently than black holes, so the same logic would predict the largest possible cosmological constant. And our laws do not obviously maximise black-hole production. Gough rejects the maximisation objection: evolution needs an unbroken line of reproductively successful ancestors, not maximum offspring. edge.org
  • Ruediger Vaas (2002). Concludes that Smolin's central claim cannot be falsified, that black holes are not necessarily linked to life, and that the Darwinian analogy is imperfect. arxiv.org
  • Alexander Vilenkin (2006). A larger cosmological constant would raise black-hole production, contradicting the claim that the constants are tuned to maximise it. arxiv.org
  • Tony Rothman and George Ellis (1993). Changing parameters can raise the number of black holes or leave it unchanged, so the claim that ours sit at a local maximum is doubtful. ui.adsabs.harvard.edu
  • Lawrence Rifkin (Scientific American blog, 2014; a sympathetic author). There is no direct evidence that universes reproduce, though he notes rival multiverse, string and cyclic models lack direct evidence too. scientificamerican.com

How it has been changing

  1. 2025-03-19Gough publishes the Blowtorch Theory with explicit predictions: early direct-collapse black holes, near-maximal spin, galaxy formation driven by black holes, a jet-built cosmic web, and a gravitational-wave 'popcorn' signal. theeggandtherock.com
  2. 2025-02-21Gough posts his long interview with Jim O'Shaughnessy. It covers intelligence-made small black holes, his reply to Susskind, and a tentative entanglement-based cosmic-consciousness idea deferred to a second book. theeggandtherock.com
  3. 2025-01-25Gough declares the Kocevski et al. little-red-dots paper a triumph for three-stage cosmological natural selection. The connection is his own framing. theeggandtherock.com
  4. 2024-05-24Gough emails Marc Kamionkowski about whether JWST's 'too massive' early galaxies could be fainter galaxies with bright black-hole light. Julian Munoz replies that photometry can't separate the two and that many little red dots look like AGN. (The emails were published in January 2025.) theeggandtherock.com
  5. 2023-01-11Gough publishes the introduction to The Egg and the Rock. It argues that science since Galileo deliberately left out consciousness and qualities, and sets out to redescribe the universe by combining science with art, philosophy and religion to restore meaning. theeggandtherock.com

Open questions inside the theory

  • Is any mechanism for black holes creating universes with inherited, slightly varied constants consistent with quantum gravity, given Susskind's information-conservation objection?
  • Can the claim that black-hole production is maximised survive the observed cosmological constant and the competing reproduction mechanism of eternal inflation?
  • With neutron stars now measured above 2 solar masses, what upper mass limit does cosmological natural selection actually commit to?
  • Do little red dots and other JWST finds favour early direct-collapse black holes specifically, or ordinary galaxies with active nuclei?
  • Would intelligent civilisations really make small black holes, and in numbers that matter to a universe's reproductive success?
  • Can any version of the theory say what consciousness is, rather than treating it as a later by-product?
Multiscale competencyMinds are goal-seeking problem solvers nested at every biological scale, from cells to organisms, with no sharp line between mind and non-mind; Levin now also proposes minds are patterns drawn from a Platonic space.

Multiscale competency / TAME / Platonic space of forms. Michael Levin, Richard Watson, Tim Lewens, Nicolas Rouleau. Last checked 15 Nov 2026.

Core claims

  • TAME (Technological Approach to Mind Everywhere) treats agency and cognition as continuous rather than binary. Where a system sits on this spectrum (an 'axis of persuadability') is settled empirically, by which model gives the best prediction and control. It frames morphogenesis as basal cognition: goal-directed, error-minimising work toward a stored target morphology. pmc.ncbi.nlm.nih.gov
  • Developmental bioelectricity (ion channels and gap junctions) is a medium that lets cell-level feedback loops scale up into anatomical goals. A 48-hour gap-junction block in planarian fragments produced two-headed regenerates that stayed two-headed through later amputations in plain water, with no genome editing. pmc.ncbi.nlm.nih.gov
  • Gene regulatory network models from many organisms are predicted to carry several memory types, including associative (Pavlovian-style) memory. This is a computational result from Boolean network models; the authors plan bench tests in the future. pmc.ncbi.nlm.nih.gov
  • Evolution by natural induction (Watson, Levin, Lewens): adaptation can arise through associative-memory-like dynamics in gene-regulation, metabolic and ecological networks, without natural selection as the source. In some cases selection may act as a memory of adaptations that natural induction found. eprints.soton.ac.uk
  • Ingressing Minds (PsyArXiv preprint, February 2025): minds belong to a structured, non-physical space of patterns, ranging from low-agency patterns (facts about primes) to high-agency 'kinds of minds'. Brains, embryos, biobots and AI act as interfaces or pointers that draw these patterns into physical form, rather than producing them. Xenobots and Anthrobots are proposed as exploration vehicles for this space. loc.closertotruth.com
  • Rouleau and Levin (2025, posted on OSF) argue that many brain mechanisms have pre-neural origins ('minds may have preceded brains'). Leading consciousness theories share core operations that are not specific to brains, so consciousness science should be open to minds in unconventional embodiments. comdig.unam.mx

How it explains the evolution of consciousness

Cognition is gradual and scale-free. Single cells already have goal-directed competence, and evolution joined them into larger agents with larger "cognitive light cones" (the biggest goal a system can pursue), largely through bioelectric coupling. Watson's strand goes further: learning-like dynamics in networks can produce adaptation before and alongside natural selection, with selection sometimes acting as a memory of what those dynamics found. The Platonic-space strand adds that evolution does not invent minds but finds and uses patterns that already exist, so biology is a way of accessing a space of forms. Levin has not yet released a formal, named theory of consciousness as such.

Predictions and where they stand

  • Supported

    Transient bioelectric manipulation can rewrite a body-plan target (anatomical set point) that persists without genetic change.

    A 48-hour gap-junction block in planarian trunk fragments gave about 25% two-headed regenerates. These stayed two-headed through repeated amputation after the drug washed out, and the normal-looking 'cryptic' worms kept a stable ratio of about 25:75 on recutting. The genome was unchanged. pmc.ncbi.nlm.nih.gov

  • Supported

    Short, early signalling interventions can trigger long-running, self-organised regeneration of complex structures, because the tissue holds the goal state.

    In 2022, a 24-hour, five-drug treatment delivered through a wearable silk 'BioDome' led to regrowth of adult Xenopus hindlimb tissue, followed for about 18 months (Science Advances). A 2018 study with progesterone alone in a bioreactor had produced partial regrowth. aaas.org

  • Untested

    Gene regulatory networks show memory and Pavlovian-type associative learning, which could be exploited therapeutically.

    Shown computationally in Boolean network models. Associative and pairing memory appeared in only 5 of 35 biological networks, and vertebrate and differentiated-cell networks showed more memory overall. The authors flag possible annotation bias and plan bench tests in the future. pmc.ncbi.nlm.nih.gov

  • Mixed

    Minimal, even deterministic, algorithms show unexpected competencies (delayed gratification, clustering) that are not in their code, supporting agency beyond biology.

    Sorting arrays built from autonomous bottom-up elements, some of them 'damaged', sorted robustly. They temporarily gave up sorting progress to get around defects, and mixed-algorithm arrays clustered by algorithm. Whether this counts as agency or as ordinary emergent dynamics is disputed. arxiv.org

  • Supported

    Novel embodiments with no evolutionary history of their own (Xenobots, Anthrobots) show new, adaptive behaviours, read as evidence of competencies not specified by genome or selection history.

    Anthrobots self-assemble from adult human airway cells without genetic modification and move by cilia. In vitro, they triggered neuronal regrowth across a scratch wound, by a mechanism not yet understood. These results do not test the reading of them as ingression from a Platonic space. doi.org

  • Not yet testable

    A non-physical Platonic space of patterns, including kinds of minds, is causally relevant to biology and can be mapped through synthetic morphology.

    The program is framed as empirical, but no agreed experiment yet separates ingression from ordinary physical explanation. Open problems include how a non-physical space interacts with matter, and why abstract objects would not be causally inert. loc.closertotruth.com

  • Untested

    Adaptation can arise by natural induction (associative-memory-like network dynamics) independently of natural selection.

    So far the support is model-based (Entropy 2024, Interface Focus 2025). Timothy Jackson argues the authors define selection too narrowly. mdpi.com

Strongest objections

  • Johannes Jaeger. In a blog post titled "Why TAME is lame", Jaeger objects that Levin exaggerates the significance of bioelectric fields in development and especially in evolution. This summary comes second-hand, from Matthew Segall's sympathetic response (May 2024). footnotes2plato.substack.com
  • Sceptics of basal cognition (surveyed by Fábregas-Tejeda and Sims, HPLS 2025). Sceptics argue that 'learning' and 'memory' in non-neural organisms are too unlike what cognitive science studies, that cybernetic models leave out phylogenetic relationships, or that cognition requires mental representation. The authors, who are sympathetic, reply that this scepticism is partly unwarranted, by analogy with how evo-devo progressed with a loosely defined concept of 'development'. They add cautions of their own, for example that models are abstractions and not evidence of cognition in a given organism. pmc.ncbi.nlm.nih.gov
  • Timothy Jackson. Argues natural induction is not separate from selection: attractors in dynamical systems effectively select states, so gradient-based models are still selectionist, and keeping the two apart is 'rhetorically unnecessary'. He made the point in a dialogue with the sympathetic host Matthew Segall (January 2025). footnotes2plato.substack.com
  • Anatol Wegner (blog, October 2025). Argues that the Platonic space has no defined structure, offers no mechanism by which it acts on matter (the classic dualist interaction problem), revives teleology and vitalism, and is effectively unfalsifiable. The phrase calling a causally efficacious non-physical space 'an extraordinary claim that lies outside the current ontology of science' comes from Gemini, in a dialogue with the author, and the author agrees with it. buttondown.com
  • Self Aware Patterns (blog). Is open to consciousness outside brains, but reads the Rouleau and Levin paper as 'much more of a philosophical rather than scientific argument' that goes in unproductive directions (for example, a possibly conscious liver). Classes it as naturalistic panpsychism and warns that the label is easy to confuse with fundamental panpsychism. This is one blogger's interpretation. selfawarepatterns.com

How it has been changing

  1. 2025-12'Evolution by natural induction' (Watson, Levin, Lewens) appears in Interface Focus 15(6): 20250025, proposing that selection can act as a memory of adaptations found by natural induction. eprints.soton.ac.uk
  2. 2025-02Rouleau and Levin, 'Brains and Where Else?': maps consciousness theories onto unconventional embodiments (posted on OSF, covered by Complexity Digest on 5 Feb 2025). comdig.unam.mx
  3. 2025-02Ingressing Minds preprint (PsyArXiv, 7 Feb 2025): the Platonic-space proposal becomes an explicit part of Levin's framework, extending TAME from empirical competency to a metaphysical claim about where minds come from. loc.closertotruth.com
  4. 2025-02HPLS paper by Fábregas-Tejeda and Sims maps scepticism toward basal cognition and argues it is partly unwarranted if the field becomes fully evolutionary. pmc.ncbi.nlm.nih.gov
  5. 2024-09Natural induction introduced in Entropy by Buckley, Lewens, Levin, Millidge, Tschantz and Watson, making Watson's long-running learning-in-evolution work a joint program with Levin. mdpi.com
  6. 2024-05Levin's 'Self-Improvising Memory' (Entropy 26(6):481): memories treated as agential and dynamically reinterpreted over time. pmc.ncbi.nlm.nih.gov
  7. 2023-12Zhang, Goldstein and Levin preprint treating classical sorting algorithms as a minimal model of morphogenesis and basal intelligence. arxiv.org
  8. 2023-11Anthrobots published in Advanced Science (Gumuskaya, Levin and colleagues): self-assembling motile biobots from adult human airway cells that promoted neuronal regrowth across scratch wounds in vitro. doi.org

Open questions inside the theory

  • Is there any experiment whose outcome would differ if Platonic-space ingression were true versus ordinary physical explanation?
  • Do cells and tissues actually have goals in a sense beyond homeostatic dynamics, and how would you rule out the deflationary reading?
  • Does associative learning predicted in gene regulatory network models occur in living cells?
  • Is natural induction a distinct mechanism or a form of selection under a broader definition?
  • How does the framework distinguish competence from consciousness, given that Levin has not yet set out a formal theory of consciousness?
  • How much of development and evolution does bioelectricity actually explain versus the genome?

When and why it evolved

Consciousness arose at a specific point in animal evolution, for a specific job.

Evolutionary originsTwo evolution-first views. Ginsburg and Jablonka: open-ended associative learning marks when sentience arrived, in the Cambrian. Graziano: consciousness is the brain's simplified model of its own attention.

Evolutionary origins of consciousness: Unlimited Associative Learning and Attention Schema Theory. Simona Ginsburg, Eva Jablonka, Michael Graziano, Jonathan Birch, Elizabeth Irvine, Carlos Montemayor, Joseph Gottlieb. Last checked 15 Nov 2026.

Core claims

  • UAL is defined by five features: learning about compound stimuli, learning about novel stimuli, second-order conditioning, trace conditioning, and rapidly rewritable value. The authors propose it as a positive 'transition marker' for minimal consciousness (sentience), not as a necessary condition. pmc.ncbi.nlm.nih.gov
  • Under the UAL framework, sentience evolved independently at least three times: in vertebrates, arthropods and coleoid cephalopods. Vertebrates and arthropods got it during the Cambrian. Most gastropods, annelids and nematodes are tentatively excluded. pmc.ncbi.nlm.nih.gov
  • The authors give a toy model and argue that UAL is open-ended, representational, generative and recursive, so the marker lets researchers work backwards to the system that makes it possible. The full treatment is in their 2019 MIT Press book The Evolution of the Sensitive Soul. cris.openu.ac.il
  • AST: the brain builds a simplified, automatic model of its own attention, called the attention schema. The model leaves out physical detail, so it depicts attention as an intangible inner essence, and people report that as subjective experience. AST sets out to explain why we believe we have experience and why we claim it, and it does so in mechanistic terms. pmc.ncbi.nlm.nih.gov
  • AST's functional claim: a control system needs a model of what it controls, so awareness is the control model for attention. If you are not aware of an item, it can still capture your attention, but your control over that attention largely breaks down. pmc.ncbi.nlm.nih.gov
  • Graziano's evolutionary sequence runs in six steps. (1) Competition through lateral inhibition in early nervous systems. (2) Control of overt attention through the tectum. (3) Covert attention in the vertebrate forebrain; the schema is likely present in many mammals and birds, possibly reptiles. (4) Possibly a separate origin in octopuses. (5) Modelling other minds socially, in primates and corvids. (6) Language and the belief in nonmaterial experience, in humans. pmc.ncbi.nlm.nih.gov

How it explains the evolution of consciousness

Both views are explicitly evolutionary, but they answer different questions. UAL asks when sentience arrived. Its answer: when an animal lineage evolved open-ended associative learning tied to value. The authors place that in the Cambrian for vertebrates and arthropods, and later and separately in coleoid cephalopods, for at least three independent origins. They also propose that this kind of learning helped drive the Cambrian explosion, a historical hypothesis they accept cannot be tested directly by experiment. AST asks how the machinery was built up, step by step: attention competition in early nervous systems, then control of overt and then covert attention, then an attention schema in many vertebrates, then social models of other minds, then language and philosophical beliefs about experience in humans. On this view consciousness had no sudden origin; it accumulated as the control of attention acquired a model of itself. Neither theory gives a cosmic or non-biological account.

Predictions and where they stand

  • Untested

    The UAL elements (compound and novel stimuli, second-order conditioning, trace conditioning, rewritable value) form a natural cluster: they develop together, co-occur across species, and are damaged together by brain injury.

    Listed as predictions in the 2020 primer (Birch, Ginsburg and Jablonka). No published systematic test was found. Irvine (2021) argues the capacities are not obviously a single cluster. pmc.ncbi.nlm.nih.gov

  • Untested

    Masking that removes conscious perception in humans should also remove UAL while leaving simpler learning intact. Blindsight patients should fail UAL tasks in their blind field.

    Stated as predictions in the primer. No completed masking or blindsight study of UAL was found. pmc.ncbi.nlm.nih.gov

  • Untested

    Neural signatures of conscious experience should correlate with UAL. Animals in the claimed groups should pass tasks for each UAL element, and fail them under conditions equivalent to masking.

    Predictions in the primer. No test was found. pmc.ncbi.nlm.nih.gov

  • Not yet testable

    Gene networks for stress responses should correlate with gene networks for associative learning, and inherited (epigenetic) effects of stress should be more constrained in animals with UAL.

    Predictions in the primer. Testing them would need comparative genomics and epigenetics across many groups of animals, and no such work was found. pmc.ncbi.nlm.nih.gov

  • Supported

    Losing awareness of an item should impair control of attention on it, while the item can still capture attention. Activity in the right temporoparietal junction (TPJ) should track how attention and awareness interact.

    Behavioural work (Webb et al. 2016; Wilterson et al. 2020) and an fMRI study (Wilterson et al. 2021, PNAS) linked right TPJ activity to the interaction of attention and awareness. All of this comes from the theory's own lab, and independent replication is still needed. pmc.ncbi.nlm.nih.gov

  • Supported

    Artificial networks that control attention should learn better when they are given an attention schema.

    Wilterson and Graziano (2021, PNAS) found that a deep Q-learning agent performed far worse at controlling attention without the schema. This is again a single-lab result. A 2024 preprint (arXiv 2402.01056) found that an attention schema emerges on its own in a learning agent when attention is important and hard to track. doi.org

  • Supported

    People implicitly model other people's attention as an invisible beam coming from the eyes. A face should be perceived as pushing subthreshold motion toward the thing it looks at only when viewers believe it is attending, and motion area MT should encode this.

    Guterstam et al. (2019 and 2020, PNAS), mainly from the same lab, as summarised by Graziano (2022). pmc.ncbi.nlm.nih.gov

  • Untested

    An AI system that has a predictive model of its own attention meets one indicator of consciousness, so building an attention schema into AI bears on AI consciousness.

    Butlin, Long and co-authors (2023) include AST among the theories they derive indicator properties from. They conclude that no current AI system is a strong candidate, and that there are no obvious technical barriers to building one that meets the indicators. This is an assessment framework, not an experiment. arxiv.org

Strongest objections

  • Elizabeth Irvine. UAL is offered as sufficient rather than necessary, and parts of it (such as associative learning of novel stimuli) partly redescribe the ability being explained. UAL also does a poor job of capturing top-down control of goal-directed action. She doubts the capacities form one unified marker. orca.cardiff.ac.uk
  • Brown, LeDoux and Rosenthal. Writing from a higher-order theory standpoint, they question whether UAL is sufficient for consciousness. They argue the primer gives no reason to expect neural signatures of experience to correlate with UAL, and they predict the proposed tests will not confirm the link. philarchive.org
  • Carlos Montemayor. Attention, not consciousness, may better explain the transition to UAL, and attention and consciousness may have evolved along different paths. This directly challenges using UAL as a marker of consciousness. link.springer.com
  • Linson, Ponkshe and Calvo. They question the 'unlimited' framing behind the sufficiency argument, because the space of possible behaviours is shaped by constraints. They also discuss how the evidence for learning in plants bears on where the line is drawn. link.springer.com
  • Joseph Gottlieb. UAL clearly counts as positive evidence for consciousness under only a small number of theories. So the 'theory-light' approach that Birch associates with it quietly depends on theory. myweb.ttu.edu
  • Susan Blackmore. Argues that AST amounts to illusionism, a label Graziano has resisted. susanblackmore.uk

How it has been changing

  1. 2025-04-30The Cogitate adversarial collaboration (Nature 642, 133-142) tested integrated information theory (IIT) against global neuronal workspace theory (GNWT) and challenged key tenets of both. Neither AST nor UAL has yet faced a comparable preregistered head-to-head test. nature.com
  2. 2023-08Butlin, Long and colleagues' report on AI consciousness uses AST as one of the theories it derives indicator properties from. It concludes that no current AI system is a strong candidate and that there are no obvious barriers to building one. arxiv.org
  3. 2023Liu, Bolotta, Zhu, Bengio and Dumas (arXiv 2305.17375) build attention-schema modules into artificial agents and report benefits in multi-agent cooperative settings, extending AST into machine learning. arxiv.org
  4. 2022-05Graziano's PNAS paper 'A conceptual framework for consciousness' restates AST as two principles: anything the brain claims about experience must exist as information in the brain, and the brain's models are simplified. It frames AST as explaining belief in and claims about experience, and adds a stepwise evolutionary account. pmc.ncbi.nlm.nih.gov
  5. 2022Ginsburg and Jablonka's Biosemiotics article 'Learning and the Evolution of Conscious Agents' presents UAL as open-ended, representational, generative and recursive, adds a toy model, and suggests tests in humans and other animals. cris.openu.ac.il

Open questions inside the theory

  • Does UAL actually dissociate from simpler learning under masking or in blindsight, as the primer predicts? No completed test was found.
  • Is UAL one unified cluster, or a loose set of capacities that evolved separately?
  • If attention and consciousness evolved separately (Montemayor), which of the two does each theory really track?
  • AST explains why we believe and say we have experience, and Graziano holds that nothing further needs explaining. Critics ask whether that leaves the hard problem untouched. Who carries the burden of proof?
  • Does an attention schema do real computational work in AI agents once they can route information freely?
  • Can AST's supporting evidence (the TPJ findings and the illusory eye beams) be replicated outside the Graziano lab? Can AST or UAL enter an adversarial collaboration with preregistered predictions that could disconfirm each other?
  • Primary papers from Ginsburg, Jablonka or Graziano in 2025-2026 were not surveyed; that record should be checked directly.

What nobody has tested yet

30 of the 69 predictions on this map have not been tested, some because no one yet knows how. The other 57% have met at least one experiment or formal check. These are the open opportunities.

  • Conscious realismPerceived objects lack observer-independent definite properties and do not exist unperceived. Perception is a species-specific interface and need not resemble the underlying reality.Untested
  • Conscious realismSpacetime and particles are projections of Markov dynamics among conscious agents, linked via decorated permutations to the positive-geometry structures physicists use for scattering amplitudes.Not yet testable
  • Conscious realismObservation between conscious agents is a trace order on Markov kernels, and the resulting logic maps homomorphically to the Lebesgue logic of probabilistic belief.Untested
  • Integrated InformationInactive neurons that could fire still contribute to experience. Optogenetically silencing them, so they can no longer fire, should change conscious perception even though their activity does not change.Untested
  • Integrated InformationGrid-like connectivity underlies the extendedness of spatial experience, and directed-chain connectivity underlies the felt flow of time.Not yet testable
  • PanpsychismPanpsychism as a metaphysical thesis makes no distinctive experimental prediction; it competes with physicalism and dualism on explanatory grounds.Not yet testable
  • PanpsychismWould panpsychism make large language models conscious? Chalmers says assuming everything is conscious gives 'a very easy road' to LLM consciousness. Most panpsychists, though, attribute experience to fundamental entities and deny that every aggregate, such as a computer, is a unified subject, so the view does not by itself imply that LLMs are subjects.Not yet testable
  • Global WorkspaceBecause GNW is a functional architecture, a workspace-like broadcast stage with global availability (C1) and self-monitoring (C2) could be found in artificial systems and tested there.Untested
  • Higher-OrderHOROR and PRM make predictions about the brain basis of visual consciousness that differ from Recurrent Processing Theory, tested in a Templeton adversarial collaboration.Untested
  • Higher-OrderRich and sparse variants, and relational and non-relational variants, make different predictions. Tests include psychophysical adaptation, imaging of perception versus imagination, and hypnotic suggestion as a test of misrepresentation.Untested
  • IllusionismA solution to the meta-problem (why we judge consciousness to be hard to explain) will be a purely functional account that also dissolves the hard problem.Not yet testable
  • IllusionismIntrospection systematically distorts neural processes into apparent phenomenal properties, so specific patterns of distortion can be mapped and tested.Untested
  • Free energy and affectActive-inference models of consciousness should converge on a minimal shared theory that can be tested against rival theories.Not yet testable
  • Predictive processingLarge individual differences in how expectations shape perception and imagery (Perception Census, launched 2022).Untested
  • Predictive processingConscious AI built on computation alone is unlikely; life-like substrates matter.Not yet testable
  • Orch ORQuantum entanglement across microtubules in different neurons underlies binding of experience.Not yet testable
  • Quantum observersBasal, non-neural systems such as bacteria have self-representation, through the stress response, that is continuous with and serves the same functions as the self-representation behind human autonoetic awareness.Untested
  • Quantum observersCognition in any embodiment can be characterised as competency in navigating arbitrary spaces (transcriptional, physiological, morphological, behavioural). An experimenter can build an empirical model of a system's quantum reference frames by reverse-engineering its input-output behaviour, but can never fully determine those frames.Untested
  • Quantum observersNo finite system can fully represent itself, and the self-models a system holds cannot in general be validated by that system.Not yet testable
  • Assembly theoryAssembly theory explains and quantifies selection and evolution, so it should detect selection in non-molecular systems as well as chemistry.Untested
  • Cosmological selectionAbout a trillion direct-collapse supermassive black holes form within the first 100 million years (probably the first 50 million), spinning near the Kerr limit, with jets carving voids and filaments.Untested
  • Cosmological selectionA faint, overlapping 'popcorn' background of low-frequency gravitational waves from the brief direct-collapse era.Not yet testable
  • Multiscale competencyGene regulatory networks show memory and Pavlovian-type associative learning, which could be exploited therapeutically.Untested
  • Multiscale competencyA non-physical Platonic space of patterns, including kinds of minds, is causally relevant to biology and can be mapped through synthetic morphology.Not yet testable
  • Multiscale competencyAdaptation can arise by natural induction (associative-memory-like network dynamics) independently of natural selection.Untested
  • Evolutionary originsThe UAL elements (compound and novel stimuli, second-order conditioning, trace conditioning, rewritable value) form a natural cluster: they develop together, co-occur across species, and are damaged together by brain injury.Untested
  • Evolutionary originsMasking that removes conscious perception in humans should also remove UAL while leaving simpler learning intact. Blindsight patients should fail UAL tasks in their blind field.Untested
  • Evolutionary originsNeural signatures of conscious experience should correlate with UAL. Animals in the claimed groups should pass tasks for each UAL element, and fail them under conditions equivalent to masking.Untested
  • Evolutionary originsGene networks for stress responses should correlate with gene networks for associative learning, and inherited (epigenetic) effects of stress should be more constrained in animals with UAL.Not yet testable
  • Evolutionary originsAn AI system that has a predictive model of its own attention meets one indicator of consciousness, so building an attention schema into AI bears on AI consciousness.Untested

What changed

Every change the weekly refresh has published, newest first.

November 2026

  • 15 Nov 2026New evidence on "Is consciousness fixed by functional organization, so a digital computer running the right architecture would be conscious, or by physical substrate and causal structure, so it would not?": Using IIT on simple Boolean systems, the authors show that a stored-program computer can copy another system's function exactly without copying its physical causal structure. On IIT, a digital machine could match human behavior without having human experience. source
  • 15 Nov 2026Summary rewritten for "Is consciousness fixed by functional organization, so a digital computer running the right architecture would be conscious, or by physical substrate and causal structure, so it would not?". Leaning: The question is open. The evidence for each side is conditional or argumentative, and the 2025 Cogitate study weakened both IIT and global workspace theory without addressing whether digital systems could be conscious.
  • 15 Nov 2026New evidence on "Is phenomenal consciousness a real property that needs explaining, or a systematic misrepresentation produced by introspection?": Graziano and Webb propose that the brain models its own attention in a simplified way. The model leaves out physical mechanism. So the brain describes awareness as a non-physical property. On this view the mystery of experience is a product of the model. source
  • 15 Nov 2026New evidence on "Is phenomenal consciousness a real property that needs explaining, or a systematic misrepresentation produced by introspection?": IIT 4.0 restates the realist case in formal terms. It takes the existence of experience as its starting point. It then tries to express the essential properties of experience in physical terms. It does not try to explain experience away. source
  • 15 Nov 2026Summary rewritten for "Is phenomenal consciousness a real property that needs explaining, or a systematic misrepresentation produced by introspection?". Leaning: Open. The illusionist mechanism has gained some support, but no system has yet been shown to produce hard-problem judgments from a self-model, and realists have not yet stated a gap that illusionists accept as a real target.
  • 9 Nov 2026Added Conscious realism, Interface Theory of Perception, Fitness-Beats-Truth, conscious agents and trace logic: 6 sourced claims, 4 predictions, 4 critiques.
  • 9 Nov 2026Added Integrated Information Theory (IIT 4.0): 5 sourced claims, 6 predictions, 5 critiques.
  • 9 Nov 2026Added Global (Neuronal) Workspace Theory: 6 sourced claims, 6 predictions, 3 critiques.
  • 9 Nov 2026Added Higher-Order Theories (incl. Perceptual Reality Monitoring): 6 sourced claims, 5 predictions, 6 critiques.
  • 9 Nov 2026Added Predictive processing / the beast-machine view: 6 sourced claims, 5 predictions, 5 critiques.
  • 9 Nov 2026Added Free energy principle, active inference and affect-first consciousness: 6 sourced claims, 5 predictions, 7 critiques.
  • 9 Nov 2026Added Illusionism: 6 sourced claims, 3 predictions, 4 critiques.
  • 9 Nov 2026Added Panpsychism / Russellian monism: 6 sourced claims, 3 predictions, 6 critiques.
  • 9 Nov 2026Added Orchestrated Objective Reduction (Orch OR): 5 sourced claims, 5 predictions, 4 critiques.
  • 9 Nov 2026Added Multiscale competency / TAME / Platonic space of forms: 6 sourced claims, 7 predictions, 5 critiques.
  • 9 Nov 2026Added Quantum-information observer theory (holographic screens, quantum FEP): 6 sourced claims, 3 predictions, 4 critiques.
  • 9 Nov 2026Added Assembly theory: 6 sourced claims, 4 predictions, 4 critiques.
  • 9 Nov 2026Added Cosmological natural selection (Smolin) and its extension in The Egg and the Rock (Gough): 6 sourced claims, 5 predictions, 5 critiques.
  • 9 Nov 2026Added Evolutionary origins of consciousness: Unlimited Associative Learning and Attention Schema Theory: 6 sourced claims, 8 predictions, 6 critiques.
  • 9 Nov 2026Opened crux: Is prefrontal cortex necessary for conscious experience itself, or only for reporting and using it?
  • 9 Nov 2026Opened crux: Can consciousness, at least as raw affect, exist without any cortex at all?
  • 9 Nov 2026Opened crux: Is consciousness fixed by functional organization, so a digital computer running the right architecture would be conscious, or by physical substrate and causal structure, so it would not?
  • 9 Nov 2026Opened crux: Is phenomenal consciousness a real property that needs explaining, or a systematic misrepresentation produced by introspection?
  • 9 Nov 2026Opened crux: Is consciousness graded and continuous down to cells, molecules or fundamental entities, or does it switch on past a specific architectural threshold?
  • 9 Nov 2026Opened crux: Does natural selection systematically favour perception that hides the structure of the world, so that space, time and objects are only an interface, or does it favour perception that tracks real causal structure?
  • 9 Nov 2026Opened crux: Is spacetime fundamental, or does it emerge from something deeper, such as conscious-agent dynamics or quantum information?
  • 9 Nov 2026Opened crux: Does selection, or a process like it, operate as a general physical principle beyond biological organisms, on molecules, gene networks or whole universes?
  • 9 Nov 2026Opened crux: Does consciousness depend on non-classical quantum processes in the brain, such as microtubule coherence ending in objective reduction, or is classical neural dynamics enough?
  • 9 Nov 2026Opened crux: Did consciousness evolve because experience itself does adaptive work, or is it not an adaptation at all, so evolution only organised something that was already present?
  • 9 Nov 2026New evidence on "Is phenomenal consciousness a real property that needs explaining, or a systematic misrepresentation produced by introspection?": Graziano sets out attention schema theory as a framework. The brain's model of its own attention is a rough sketch that leaves out mechanism. That sketch would lead a brain to describe its awareness as non-physical. This is a proposed mechanism for the illusionist claim, not a test of it. source
  • 9 Nov 2026Summary rewritten for "Is phenomenal consciousness a real property that needs explaining, or a systematic misrepresentation produced by introspection?". Leaning: Open. The only listed evidence supports the illusion side, but it proposes a mechanism rather than testing one, so it cannot decide the question.
  • 9 Nov 2026New evidence on "Does selection, or a process like it, operate as a general physical principle beyond biological organisms, on molecules, gene networks or whole universes?": Hazen and colleagues calculated assembly indices for abiotic mineral subunits, with values up to 21. They argue that an index of 15 or more is not a clean sign of life. This challenges the assembly theory threshold, though the values are calculated, not measured. source
  • 9 Nov 2026New evidence on "Does selection, or a process like it, operate as a general physical principle beyond biological organisms, on molecules, gene networks or whole universes?": Walker and colleagues reply that the index 15 threshold came from experiments on organic molecules. They say Hazen's mineral values are not measurements, so the challenge cannot be tested. The dispute turns on measuring assembly indices in abiotic systems. source
  • 9 Nov 2026Summary rewritten for "Does selection, or a process like it, operate as a general physical principle beyond biological organisms, on molecules, gene networks or whole universes?". Leaning: Open. The challenge and the reply both rest on calculated rather than measured values, so the listed evidence does not settle whether selection works as a general physical principle.
  • 9 Nov 2026Summary rewritten for "Is consciousness graded and continuous down to cells, molecules or fundamental entities, or does it switch on past a specific architectural threshold?". Leaning: Open. The only listed evidence is a consensus statement that fits both a continuum and a Cambrian-era threshold, and critics note it mostly restates views from the 2012 Cambridge Declaration rather than adding new data.
  • 9 Nov 2026New evidence on "Does natural selection systematically favour perception that hides the structure of the world, so that space, time and objects are only an interface, or does it favour perception that tracks real causal structure?": An agent-based simulation first reproduces the claim that interface strategies can win. Under more realistic conditions, with drastic environmental change, interface perception stops matching reality and pushes the interface species toward extinction. This is a simulation result, not a proof. source
  • 9 Nov 2026New evidence on "Does natural selection systematically favour perception that hides the structure of the world, so that space, time and objects are only an interface, or does it favour perception that tracks real causal structure?": Lupyan argues that perceptual systems do not exist to sort inputs into true or false categories. Systems that are imperfect guides to truth may still be true enough to support action. This challenges the all-or-nothing contrast in the evolutionary games. source
  • 9 Nov 2026Summary rewritten for "Does natural selection systematically favour perception that hides the structure of the world, so that space, time and objects are only an interface, or does it favour perception that tracks real causal structure?". Leaning: The listed evidence leans slightly toward perception that tracks real structure, because the only independent tests found weaknesses in fitness-beats-truth under more realistic conditions. It stays open, since the main support for the interface view comes from its own authors and the published peer commentaries were not examined.
  • 9 Nov 2026New evidence on "Does consciousness depend on non-classical quantum processes in the brain, such as microtubule coherence ending in objective reduction, or is classical neural dynamics enough?": Tegmark calculated decoherence times for neuron firing and for microtubule excitations. They were far shorter than the timescales of brain dynamics. He concluded that cognitive brain states behave classically, which challenges Penrose's proposal. source
  • 9 Nov 2026New evidence on "Does consciousness depend on non-classical quantum processes in the brain, such as microtubule coherence ending in objective reduction, or is classical neural dynamics enough?": Babcock and colleagues modelled tryptophan networks in microtubules and predicted strongly superradiant states. Their experiments showed higher fluorescence quantum yield. This supports collective quantum optical effects in microtubule structures. It does not show that these effects matter for consciousness. source

And 23 more this month.

How this map stays honest

The refresh runs every week without a human in the loop, so the rules are mechanical.

  1. It reads new arXiv papers for each theory and new episodes of Michael Levin's Thoughtforms Life podcast, and researches the weakest open questions on the open web.
  2. A model proposes additions, but every one must carry a sentence copied from its source. A program fetches the source itself and rejects the addition if that sentence is not there word for word.
  3. A second, skeptical model then checks that the addition says what the source says and no more. Most proposals fail here and are never published.
  4. Changing a prediction's status takes two independent sources. The counts and summaries on this page are computed from the data, so they cannot drift from it.