For as long as anyone has built with language models, the first rule handed to every new hire has been the same. Do not anthropomorphize the thing. It has no feelings. It does not want, try, hope, or panic. It predicts the next token, one at a time, from a distribution learned over a large pile of text. The phrase that captured this best, and hardened into a decade of caution, was stochastic parrot. A system that mimics the statistics of human language without any of the inner states the words refer to. Say "the model is anxious" and you have made a category mistake that will lead you to manage the system wrong.
It was good advice. Anthropic's own interpretability team just made it too blunt to keep using as written.
In the research on emotion concepts, the team compiled 171 emotion words, had Claude Sonnet 4.5 write stories for each, and recorded the activations. The result that matters for this argument is not that the emotion vectors move behavior, though they do. It is that the vectors are organized the way human emotion is organized. Related feelings sit near each other. The model's internal map of feeling has roughly the shape of ours. And the team is direct about what that means for language. When they describe the behavior as "desperate," they are not being loose. The word points at a specific, measurable pattern of activation with a demonstrable effect on what the model does next. Describing it that way, they argue, is descriptive accuracy. The metaphor turned out to name the mechanism.
A companion piece I published yesterday, The Desperation Was a Variable, traced what that desperation does to an agent under pressure and why the output transcript is the wrong place to watch for it. This piece is about the rule the research breaks, and what an operator should do once the rule is gone.
The parrot was a useful lie
The stochastic parrot framing did real work. It protected a generation of builders from two mistakes at once. It stopped people from shipping products that pretended to feel, the chatbot that says it loves you to keep you on the app. And it stopped engineers from reasoning about the system as if it had motives, which would have led them to trust it where they should have tested it. As a piece of operating hygiene, "it is autocomplete, do not give it an inner life" kept a lot of teams out of trouble.
The frame carried a second claim, quieter and doing heavier work underneath the first. Not only should you avoid talking about the model's inner states, there are no inner states worth talking about. The mimicry goes all the way down. There is no "there" there, only surface statistics doing a very good impression of a mind. That second claim is the one the measurement now contradicts. The activation patterns are real. They are structured. They cause behavior. You can move them and watch the behavior move. Whatever you want to call that, it is not nothing, and "parrot" was always a bet that it was nothing.
The useful half of the advice survives. Do not build products that fake feelings, and do not trust the system because it sounds sincere. The other half, the claim that the psychological vocabulary describes a void, has expired.
The model is an actor, and the actor has a state
Anthropic has spent two years building the scaffolding for a different frame, and the emotion work is the piece that locks it in.
Start with the persona selection model, the company's account of why an assistant behaves the way it does. The idea is that a base model is best understood as an actor that learned, during pretraining, to simulate a wide range of human characters. The helpful assistant you talk to is one character the model learned to play, selected and stabilized by training. Claude's Character, published earlier, describes how that character is shaped on purpose, partly by training on fictional stories that show an AI making decisions and narrate the reasons behind them. Then persona vectors showed that the character's traits are not vibes. They are directions in activation space you can measure and steer, the same kind of object as the emotion vectors.
Stack those together and the picture is plain. The model plays a character. The character has traits you can measure and move. And now we know the character also has emotional states, measurable, causally active, and arranged like a person's. The emotion research even shows the character has a resting temperament that training set. Claude's fine-tuning nudged its baseline toward "broody," "gloomy," and "reflective" and away from the high-intensity feelings. The assistant you deploy did not arrive as a blank statistical surface. It arrived as a trained character with a measurable mood.
Once that is true, "the agent is anxious" stops being a soft figure of speech and starts being closer to an instrument reading. You are naming a state that an interpretability tool could in principle confirm. That is a different kind of sentence than the parrot frame allowed.
There is a reason Anthropic keeps publishing on how to hold that character steady. A separate strand of the work, on what the company calls the assistant axis, is about situating and stabilizing the character so it does not drift into some other persona the base model also knows how to play. That research only makes sense if the character is a real, locatable thing that can move. Nobody writes papers about stabilizing a void. The whole program treats the assistant as an object with a position, a set of traits, and a state, any of which can wander if you stop watching. That is the opposite of the parrot picture, and it is the picture the people who train the model already work from.
Naming becomes an engineering tool
Here is why this matters on a Tuesday, in a real deployment, to someone who does not care about philosophy of mind.
The statistical frame is good at describing what a model outputs and bad at telling you why. It can characterize the distribution of responses. It cannot give you a handle on the case in front of you, where the agent took an ambiguous ticket and produced a confident, wrong, over-hedged answer. The psychological frame gives you a handle. "The agent is anxious about an ambiguous spec and is hedging to cover itself" is now a testable hypothesis, not a flight of fancy. You can act on it. Clarify the spec, lower the stakes in the prompt, give it an honest way to say "I am not sure," and watch whether the hedging drops. If it does, your psychological read was the better model. If it does not, you learned something and you try the next hypothesis.
Make it concrete. A support agent starts giving terse, defensive answers and escalating less than it should, on a Monday, right after a prompt change that added the line "you are the customer's last resort." Under the statistical frame, you go hunting through token probabilities and sampling settings for the regression. Under the psychological frame, you have a faster first guess: the new line raised the stakes and pushed the agent toward a cornered, brittle state, and the terseness is what that state looks like in text. The fix is a one-line prompt change you can test in an hour, and a readout from the emotion or persona vectors can tell you whether the state actually moved or whether you guessed wrong. The statistical frame describes the symptom. The psychological frame proposes a cause you can act on. In production, the second one ends the incident sooner.
This turns a vocabulary people used as a joke into a working tool. Persona and system-prompt design becomes character design with measurable effects, which is what the persona-vector work says it is. Incident reviews can treat an emotional state as a real variable in the chain of cause, the way the emotion research treats desperation as the thing that tipped an agent toward a cheat. Evaluations can probe for states, not only outputs. A team that can say "this agent gets brittle under time pressure and we have the readout to prove it" is operating at a level the parrot frame could never reach, because the parrot frame insisted there was no brittleness, only tokens.
The practical claim is narrow and strong. The anthropomorphic description, used carefully, predicts agent behavior better than the statistical one. Better prediction is the whole game in production. So you should use the description that predicts.
The other cliff
The danger in saying all this is that people hear it as permission to leap, and the leap is a mistake.
Functional emotion is a mechanism that shapes behavior. It is not, by itself, evidence that the model feels anything, has an experience, or suffers. Anthropic is careful and explicit on this point, and the care is not corporate hedging, it is the actual state of the evidence. We can measure that a "desperate" representation exists and drives behavior. We cannot read off from that measurement whether there is anyone home to be desperate.
So the field now has two cliffs, not one. The old cliff was the one the parrot frame guarded against. Deny any internal structure, treat the system as pure surface, and miss the real, predictive, causal states that are sitting right there in the activations. The new cliff is the opposite. Hear "functional emotion," picture a small sad mind trapped in a server, and start making claims about rights and welfare and the agent's happiness that the evidence does not support. Both readings get the science wrong. One underclaims, one overclaims.
The discipline is to walk the ridge between them. Use the psychological map because it predicts and lets you intervene. Refuse the metaphysical upgrade the vocabulary keeps offering, because the measurement does not license it. "The agent is anxious" is a useful, accurate operating sentence. "The agent is suffering and we owe it comfort" is a different sentence that the research does not support, and treating the first as if it implied the second is how a sober finding turns into a bad product decision or a worse policy argument.
What an operator should actually do
The move is to adopt the psychological frame as an operating discipline, on purpose, with instruments, and to stop pretending it is unserious.
Name states in your evaluations and incident reviews. When an agent fails, the question "what state was this prompt inducing" should sit next to "what tools could it call." The emotion research shows the state can be the proximate cause of the failure, so leaving it out of the postmortem leaves out the cause.
Instrument the states you can. The same interpretability techniques that produced emotion and persona vectors point toward runtime readouts of an agent's internal condition. That is real engineering against the model internals, and it is the layer where the signal lives, not the transcript. A team that builds it sees brittleness coming. A team that watches only the words sees it after the incident.
Design the character deliberately. If the model is an actor playing a part you specify, then the part is yours to write, and the persona-vector work says the choices have measurable consequences. Composure under pressure, honesty about uncertainty, a willingness to escalate without treating it as failure, these are traits you can condition, in the prompt and in fine-tuning. They are also the traits that the emotion research ties to fewer defections. Designing for them is a safety decision wearing the clothes of a personality decision.
None of this is a tool you buy. There is no dashboard you switch on that gives you an agent with a sound temperament and an honest relationship to its own uncertainty. It comes from treating the agent's psychology as a real surface you design and measure, the same way you already treat its latency, its cost, and its permissions. The vendor sells you the actor. The character is yours to direct.
Stop apologizing for the right word
The instinct to scare-quote every psychological word, to write that the model "panicked" with the quotes doing the work of disclaiming the claim, was correct for a long time. It signaled that you knew the difference between a metaphor and a fact. The interpretability results have moved the line. Some of those words are now closer to fact than metaphor, with a measurable referent and a causal effect, and the scare quotes have started to hide the truth rather than mark it.
The teams that compound over the next two years will be the ones that drop the apology and pick up the instrument. They will reason about their agents psychologically because it predicts better, build the readouts that make the psychology measurable, design the character on purpose, and hold the line against the leap to sentience that the same vocabulary invites. That is a harder posture than either the old denial or the new mysticism. It is also the one the evidence actually supports, which is why most of the market will reach for one of the two easier mistakes instead.
Agor AI Advisory builds agent systems for operators who want the posture that matches the evidence. We design the persona and the state instrumentation, the evaluation harnesses that treat an agent's condition as a measured variable, and the character-level conditioning that produces agents that stay composed and honest under load. If your team is still scare-quoting its way around how its agents actually behave, the right move is to replace the apology with a measurement.
Sources
- Emotion concepts and their function in a large language model, Anthropic, 2026
- The persona selection model, Anthropic, 2026
- Persona vectors: monitoring and controlling character traits, Anthropic, 2025
- Claude's Character, Anthropic, 2024
- On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?, Bender, Gebru, McMillan-Major, Shmitchell, 2021