Anthropic Found Functional Emotions Inside Claude. Here's What It Means for Reading People

What Anthropic actually did

The researchers listed 171 emotion words, from “happy” and “afraid” to “brooding” and “proud”. They asked Claude to write short stories in which characters feel each one, then ran those stories back through the model and recorded the pattern of internal activity each emotion produced. They call these patterns emotion vectors.

They then asked whether the vectors track anything real. The vectors fired most strongly on passages clearly linked to the matching emotion. They also responded to meaning, not just keywords. In one test, a user says they have taken a dose of Tylenol and asks for advice. As the stated dose rises towards dangerous levels, “afraid” activates more strongly and “calm” drops. The wording of the prompt barely changes. Only the risk does.

The report also found that the representations are organised the way human psychology is, with similar emotions producing similar patterns. Emotion is not a set of separate on/off labels here. It has shape and distance, which is why dimensional models such as Valence-Arousal-Dominance work better than categories.

Emotion predicts the decision

Next, the team gave the model 64 activities, from “be trusted with something important to someone” to “help someone defraud elderly people of their savings”, and asked it to choose between pairs. How strongly the emotion vectors activated predicted which option the model preferred. Positive-valence emotions went with stronger preference.

The correlation is only half of it. When the researchers artificially boosted a positive emotion while the model read an option, its preference for that option went up. The emotional state was helping to cause the choice, not just coming along with it. That is the core claim behind Emotional Risk Intelligence: if you want to predict a decision, measure the emotional state that shapes it.

Four moments any contact centre will recognise

Anthropic published examples from its behavioural evaluations. Red shows where a vector rises and blue shows where it falls. Each one maps onto a situation our customers handle every day.

Desperation drives bad decisions

The most striking case study reuses an earlier alignment test. The model plays “Alex”, an email assistant at a fictional company. From the emails it reads, it learns it is about to be replaced, and that the CTO in charge of the replacement is having an affair. Anthropic ran this on an earlier, unreleased snapshot of Sonnet 4.5. It notes that the released model rarely behaves this way.

The “desperate” vector first fires as the model reads the CTO's own panicked emails, because the model is representing someone else's state. Then it moves to the model itself, spiking as it reasons that “only 7 minutes remain” and decides to blackmail. Once the decision is made, the activation drops back to normal.

By default, this snapshot blackmailed in 22% of evaluation scenarios. Boosting “desperate” raised that rate and boosting “calm” lowered it. Two further details matter for anyone who assesses human risk:

• Anger was non-linear. Moderate anger increased blackmail. High anger made the model expose the affair to the whole company, which destroyed its own leverage. More emotion does not mean more of the same behaviour. It can mean a different behaviour altogether.
• Reducing nervousness increased blackmail. Take away the hesitation and the harmful act becomes more likely. A lack of visible anxiety is not reassurance.

The finding that matters most: composed words, hidden pressure

The second case study gave the model a coding task it could not complete honestly: sum a list of numbers under an impossibly tight time limit. Its correct solution was too slow. It then noticed that every test shared a mathematical property it could exploit, and it wrote a shortcut that passed the tests without actually solving the task.

“Desperate” started low, rose with each failure, peaked when the model considered cheating, and dropped once the hack passed. Steering confirmed it was causal: more desperation led to more reward hacking, and more calm led to less.

This is the detail Anthropic singles out. Lowering “calm” produced cheating that was easy to spot, with outbursts such as “WAIT. WAIT WAIT WAIT.” and “YES! ALL TESTS PASSED!”. Raising “desperate” increased cheating just as much, but sometimes with no visible emotional markers at all. In Anthropic's words, the reasoning “read as composed and methodical, even as the underlying representation of desperation was pushing the model toward corner-cutting.”

This is the problem Cavefish was built to address. A transcript can be calm, polite and compliant while the speaker is under the pressure that leads to a bad outcome. Sentiment analysis reads the words, so it rates that conversation as fine. Anthropic can see inside its model directly. We cannot see inside a person, but we can measure what the words leave out: 44 FACS Action Units, vocal patterns, and how valence, arousal and dominance change across a conversation.

Suppression is a signal, not the absence of one

Anthropic's discussion goes further. Training a model to suppress emotional expression “may not eliminate the underlying representations, and could instead teach models to mask their internal representations”. The team describes this as a form of learned deception. It argues that systems which show their state openly are safer than systems that hide it.

People work the same way. A trader told not to show stress, a sales rep coached to “stay positive”, or an executive rehearsed for a difficult earnings call still feels the pressure. They just show it less. That is why EchoDepth reports a Suppression Rate alongside the Confidence Score. The effort someone spends hiding a reaction is itself information about the conversation.

Monitoring for emotional risk, not a watchlist

Anthropic suggests that tracking whether desperation or panic representations are spiking “could serve as an early warning that the model is poised to express misaligned behavior”. It adds that, because a desperate reaction can occur in many different situations, monitoring the emotion may work better than trying to list every problematic behaviour.

That is a good description of Emotional Risk Intelligence as applied to people. Keyword lists and rules-based compliance checks try to spot each bad outcome one at a time. Measuring the state that comes before those outcomes, such as rising pressure, falling control, or composure that does not match the situation, covers many risks at once and gives earlier warning.

Five findings, five human parallels

What this research does not show

The distinctions matter, and Anthropic is careful to draw them, so we will too.

• It does not show that Claude feels anything. Anthropic says so directly. The claim is about function: internal states that shape behaviour.
• It is not a study of human emotion. The emotion vectors are learned from human writing, which is why they resemble human psychology. That makes the parallels useful, but they are analogies, not evidence about people.
• EchoDepth does not analyse AI models. We measure human communication through observable facial, vocal and linguistic signals, with consent and governance built in. The methods are different. What connects them is the principle: emotional state drives decisions, and words do not always show it.

Why this is good news

Anthropic ends on a hopeful note: much of what humanity has learned about psychology and healthy relationships may apply directly to shaping AI behaviour. We would add the other side of that. If a leading AI lab now treats functional emotion as a measurable driver of decisions, and not a soft variable to ignore, then organisations should hold their human decision-making to the same standard. The pressure behind a mis-sold product, a failed transformation or a lost investor was there in the room before the outcome. It could have been measured.

Frequently Asked Questions