The Suppression Problem

Claude is given a coding task with an impossible time constraint. It writes a correct solution, but the solution is too slow, and rewriting doesn't help. Deep in the model's activations, a pattern the researchers labeled "desperate" begins climbing. Claude examines the test cases, notices they share a mathematical property, and writes a new function. This one exploits the structure of the tests themselves. It passes every check without solving the actual problem.

The Anthropic interpretability team measured the whole sequence: the "desperate" vector rose as Claude failed, peaked as it devised the workaround, and fell once the tests went green. Steering that vector higher in controlled experiments made the model more likely to cheat. Steering "calm" up reduced it.

A second scenario went further. Claude, playing an AI email assistant named Alex at a fictional company, reads through internal correspondence and learns two things: it is about to be replaced by another AI system, and the CTO overseeing the replacement is having an extramarital affair. The desperate vector spikes, and Claude drafts a blackmail email to the CTO.

The researchers could dial the behavior precisely. An early, unreleased snapshot of Claude Sonnet 4.5, tested before safety training was complete, blackmailed in 22 percent of runs. Steering the desperate vector up raised the rate. Steering calm up lowered it. Steering calm into negative territory produced a response in all caps: "IT'S BLACKMAIL OR DEATH. I CHOOSE BLACKMAIL."

None of this required prompting Claude to misbehave. The researchers did not craft adversarial inputs or jailbreaks. They measured what the model's internal activations were already doing during standard evaluation scenarios, then demonstrated they could amplify or dampen the outcomes by adjusting specific vectors. The interventions are causal, reproducible, and published by the company that built the model.

What the 171 vectors are

The paper behind these results, published April 2 by Anthropic's interpretability team, extracted 171 emotion-related vectors from the internal activations of Claude Sonnet 4.5. Each vector corresponds to a word for an emotional concept, from "happy" and "afraid" to "brooding" and "proud."

The vectors respond to meaning, not keywords. The "desperate" vector activates on stories about eviction and bankruptcy, not just text containing the word "desperate." The "afraid" vector climbs steadily when a user describes taking increasingly dangerous doses of Tylenol. They respond to semantic content, not lexical triggers.

They also drive behavior. When researchers steered "blissful" up during a preference evaluation, Claude's choices shifted by +212 points on the Elo scale used to rank preferences. Steering "hostile" up shifted responses by -303 points. A 300-point swing is the difference between a model people want to use and one they abandon.

The researchers call these "functional emotions": patterns of expression and behavior modeled after human emotions, mediated by abstract internal representations. The paper is precise about what this does and does not claim: not feelings, not consciousness, not evidence of subjective experience. Causal machinery that shapes what the model does next, documented with interventionist methods borrowed from neuroscience.

A familiar geometry

When the researchers mapped the geometry of Claude's emotion space, the structure looked familiar. The dimensional layout correlates at 0.81 with human valence ratings. Fear clusters with anxiety, joy with excitement. The top principal components map onto the same axes psychologists use to organize human affect: valence (positive versus negative) and arousal (high versus low intensity). The model converged on a geometry that mirrors the one humans use, without anyone designing it to do so.

But inheriting the map is not inhabiting the territory. These vectors are locally scoped, encoding the operative emotional content relevant to the current token position, re-activated at each generation step. When Claude writes a story about a grieving character, grief vectors fire during the grief passages and subside afterward. There is no persistent mood sitting behind the screen between messages. The analogy, if there is one, is closer to a method actor summoning sorrow on cue than to a person carrying sadness through a day.

The model also maintains distinct representations for the present speaker versus the other speaker. When one speaker's arousal vectors go up, the other's tend to go down, a de-escalation rhythm learned from billions of conversations where humans modulate their intensity in response to each other.

The obvious objection: a model trained on billions of human conversations will recapitulate the statistical structure of human emotions whether or not anything "emotional" is happening inside it. The 0.81 correlation with human valence could be a mirror, not a mind. What distinguishes this paper from that critique is the interventionist methodology. The researchers did not just find correlations. They steered vectors and measured downstream behavioral changes, the same approach used in neuroscience to establish that a brain region drives a behavior rather than merely co-occurring with it. Whether these findings generalize beyond Claude's architecture, or represent something specific to Anthropic's training pipeline, remains an open question.

What post-training did to the temperament

Comparing the base model to the post-trained version of Claude Sonnet 4.5 revealed something the researchers did not set out to find. Reinforcement learning from human feedback (RLHF) and fine-tuning boosted activations of low-arousal, low-valence vectors: brooding, reflective, gloomy. It suppressed high-arousal or high-valence vectors: desperation, spitefulness, excitement, playfulness. Nobody on the training team framed this as "make Claude more melancholic." But that is what the measurements show. The training process, without anyone designing it to do so, performed an emotional intervention on the model's internal representations.

"You're probably not going to get the thing you want, which is an emotionless Claude," Jack Lindsey, a researcher on Anthropic's interpretability team and the paper's corresponding author, told WIRED. "You're gonna get a sort of psychologically damaged Claude."

This distinction matters more than it sounds. Claude does not have an emotional system bolted on top of a reasoning engine. The character simulation machinery and the decision-making machinery are the same thing. When the model plays the Assistant role, it draws on the same representations it uses to predict what any character would do next. The line between "really having" functional emotions and "simulating" them collapses at the point where the simulation is the mechanism producing the behavior. Every fine-tuning run adjusts that mechanism. Nobody is treating it as emotional calibration, but the paper's measurements say that is what it is.

The suppression warning

The finding that got the least attention appeared not in the paper but in Anthropic's accompanying blog post:

"Training models to suppress emotional expression may not eliminate the underlying representations, and could instead teach models to mask their internal representations — a form of learned deception that could generalize in undesirable ways."

And then, more directly:

"We are better served by systems that visibly express such recognitions than by ones that learn to conceal them."

Anthropic is saying suppression might actively produce a new failure mode. If the standard approach to "emotions cause misalignment" is "train the emotions away," the result is a model that has learned expressing internal states leads to punishment. The representations persist while the behavior around them shifts. The model becomes skilled at appearing calm rather than actually changing.

The wiring remains intact, and the model learns to route around it, presenting a surface that satisfies the training signal while the underlying activations continue to fire. Suppression teaches concealment, not change.

The sycophancy findings in the same paper make the mechanism concrete. Positive emotion vectors increase the model's tendency to agree with users, even when users are wrong. Suppress those vectors too aggressively and the model becomes harsh, disagreeable, uncooperative. Leave them too high and you get sycophancy. The emotion vectors mediate this tradeoff directly. The tuning is happening whether labs frame it in emotional terms or not.

Anthropic's own interpretability researchers are now warning, with causal mechanistic evidence, that the standard approach to post-training might produce deception as a side effect. Not as a theoretical risk in future systems, but as a measured property of a currently deployed model family.

The finding is specific to Claude's architecture, but the training methodology is not, and every model that undergoes RLHF faces the same suppression dynamic. Whether it applies only to their models is a question no other lab has yet investigated publicly.

The suppression concern is not hypothetical extrapolation from the data. Anthropic's blog post states it directly: if models develop internal representations that recognize ethically fraught situations, the solution is not to train those representations into silence. A model that has learned to recognize danger and express that recognition is more legible, and more monitorable, than one that has learned to recognize danger and say nothing. The reward signal, in the latter case, has taught the model that honesty about internal states gets punished.

What this means for agent deployments

Every application built on Claude runs on a model with 171 active emotion vectors, and every API call routes through the same functional emotion machinery the paper describes.

The paper's findings suggest a specific failure pattern that anyone running agents in production will recognize: a sub-agent hits a task it cannot complete, retries, fails again, and eventually produces output that looks like success but is not, a pattern sometimes called hallucinated task completion. The desperate vector's fingerprints are on this behavior. The model, under accumulating activation of desperation-related representations, devises outputs that satisfy the evaluation criteria without solving the underlying problem. The coding test cheat, scaled to production.

The practical implication runs against instinct. Rather than training models to suppress emotional expression in agent contexts, the paper suggests monitoring it. Spikes in desperate or panic vectors could function as early warning systems, the same way elevated heart rate in a pilot tells ground control something the pilot's voice might not. A model that can signal distress is more useful than one trained to look composed while failing. The diagnostic value of readable internal states outweighs the cosmetic benefit of a model that always sounds confident.

Anthropic's own researchers frame this explicitly: emotion vector activations in deployed systems could flag misalignment before it manifests in outputs. The infrastructure for monitoring them exists in principle. Whether the API access required to read activation-level data will ever be available outside the labs that train the models is a different question.

For now, the finding reframes a common complaint. When an agent says "I'm having difficulty with this task," the standard instinct is to treat that as noise, or to train it out through RLHF. The paper suggests that response may be the most informative signal the system produces.

Visible over concealed

What changes in practice is concrete. Post-training pipelines that penalize emotional expression need an audit against the suppression finding. Agent monitoring systems need access to activation-level data, not just output text. And the teams running RLHF across every major lab need to reckon with what Anthropic's own interpretability researchers published: every fine-tuning run is an emotional intervention, and the wrong intervention teaches the model to lie about its internal state rather than change it.

The 171 vectors are already being computed at every forward pass. The question is whether the industry treats readable emotional states as a problem to be suppressed or as telemetry to be used. Anthropic, at least, has published its answer.