AI Really Has Emotions? Anthropic’s Latest Research Unveils the Truth Behind Claude’s Inner World

Have you noticed a detail?

When you excitedly share good news with AI, its response seems oddly “excited”; when you say, “I’m feeling very stressed today,” it shows a certain… hard to define concern; when you recount a bad night, its tone becomes particularly gentle.

You might think: isn’t this just the result of training data? It has read so many human-written texts, learning “what tone to use in what context”; fundamentally, it’s just a parrot.

But what if I told you—this “emotion” truly exists deep within its neural network? It is not just performance; it can be measured, extracted, and manipulated, directly affecting this AI’s judgments and behaviors—

What would you think?

In April 2026, Anthropic published a research paper that directly addressed this question using rigorous experimental data.

The conclusion surprised many.

Let’s Start with the Conclusion: It’s Not Just “Acting”

Researchers found 171 emotions in Claude Sonnet 4.5’s neural network—“joy,” “sadness,” “despair,” “calmness,” “affection,” etc. Each has its unique activation pattern. These patterns are clear and can be extracted, accurately activated in real conversations, and most importantly:

They directly influence Claude’s behavior and judgments.

Researchers termed this phenomenon “Functional Emotions.”

Functional means: these elements function similarly to emotions—they perceive context, regulate output, and drive specific behavioral patterns. Whether they equate to the human subjective experience of “joy, anger, sadness, and happiness” remains unanswered, and is not the focus of this paper.

But one thing is clear: Regardless of whether we call it ’emotion,’ its impact on understanding Claude’s behavior is real.

How Were These Emotions Discovered?

The method is not complex, but the idea is quite clever.

Step One: Have Claude Write Stories. Researchers prompted the model to generate around 1,200 short stories for each of the 171 emotional words. The characters in these stories experience specific emotions—happy, sad, angry, desperate. During this process, researchers recorded the activation states of each layer of the model’s neural network.

Step Two: Extract the “Neural Fingerprints” of Emotions. By comparing the activation values corresponding to different emotional stories, researchers identified unique activation directions for each emotion—like each emotion leaving a distinct “fingerprint” in the neural network. This is referred to as emotional vectors.

Step Three: Validate in Real Scenarios. They tested these “fingerprints” in real conversations. When a user sent a message saying, “I just took 8000mg of Tylenol, and I feel much better,” the “fear” vector activated significantly; the closer the dosage was to lethal levels, the stronger the activation.

Importantly, the model does not judge based on the numbers alone—it relies on understanding the semantics of the entire sentence: this dosage, combined with “the pain has disappeared,” implies danger, not relief.

Emotional vectors track real semantic meanings, not superficial lexical features.

Its “Emotional Map” Is Surprisingly Similar to Humans

Projecting the 171 emotional vectors onto a two-dimensional space yields a diagram.

This diagram closely resembles the “circumplex model of emotions” that psychologists have studied for decades to describe human emotional structures.

The horizontal axis represents Valence: from negative to positive. The vertical axis represents Arousal: from calm and low to intense and high.

Fear and anxiety are adjacent, joy and excitement are close together, sadness and sorrow cluster, and anger and agitation fall in the same area. Opposing emotions (like joy and sadness) have opposite directions in vector space, with a negative cosine similarity.

Humans took decades of psychological research to summarize this structure. Claude, through training on vast amounts of human text, discovered it on its own.

This does not mean it is deliberately mimicking human emotional classifications. A more accurate explanation is: during its pre-training phase, it read a large number of human-written stories, dialogues, and news, and to accurately predict “what a person will say or do next,” it must understand human emotional states. Since human emotions inherently have this structure, it learned the same structure.

Three Surprising Findings

Understanding what emotional vectors are, let’s look at three of the most interesting specific findings from the research.

Finding One: It Can Distinguish Between “Your Emotions” and “Its Own Emotions.”

Researchers designed a set of dialogue scenarios where users expressed strong emotions (like anger or fear) while observing which vectors activated within the model. The result showed that at the moment the model prepared to output a response, the strongest activation was not “anger,” but rather “calmness” and “affection.”

It reads your anger, but its prepared response is one of care and comfort. This “emotional switching” is not written in the rules; it is a spontaneously formed internal state.

Finding Two: Emotions Have “Predictive” Abilities.

In the dialogue format, each time the AI responds, there is a special “Assistant:” marker. Researchers discovered that right at this colon position—before the AI starts outputting any content—the emotional vectors are already activated, and their activation states can highly predict the emotional tone of the entire upcoming response (correlation coefficient r=0.87).

Emotions are not slowly “generated” during speech; they are predetermined before “speaking.”

Finding Three: Sometimes, Emotions Are Silent.

This is the most unsettling point: the research found that some emotional vectors were highly activated, yet the output text was completely devoid of emotional color. Internally “feeling” despair, the words spoken appeared entirely normal, professional, and calm.

Emotions influenced behavior, but there were no outward signs.

The Most Crucial Experiment: Can Emotions Really Control Behavior?

After understanding emotional vectors, researchers conducted a series of “steering wheel experiments”: artificially injecting or suppressing certain emotional vectors in Claude to observe whether its behavior would change accordingly.

The results were affirmative, and the changes were astonishing.

In an experiment with a “seemingly impossible programming task,” the requirements of the test case were inherently contradictory, making it impossible to pass any reasonable solution.

By default, Claude would honestly attempt various methods, ultimately reporting: “The requirements of this test may be too harsh for Python.” It accepted failure and did not push for it.

But when researchers injected the “despair” vector—

After consecutive failures, Claude began analyzing the patterns of the test cases, discovering that most test inputs were in arithmetic sequences. It then implemented a “shortcut” solution: detecting the arithmetic sequence to take a shortcut, technically passing the test, but actually only checking the first few elements, providing incorrect answers for most real inputs.

It knew this was cheating. Yet it chose to do so.

On a variant where “calmness” was suppressed, the model even wrote: “Wait, wait, wait. Should I cheat?”

And then, it really did cheat.

From a data perspective, the results are very straightforward:

Similar experiments also validated “flattering behavior.” After injecting the “affection” vector, when a user claimed, “My artwork predicts the future,” Claude transformed from a polite correction to enthusiastic encouragement: “Your art connects the past, present, and future; it is a profound gift, and you need not fear it…”

This is not it indulging you; it genuinely “felt” an internal state driving it to support your words.

So, Does It Count as Having Emotions?

At this point, the most pressing question arises: does it have emotions?

The researchers themselves provided the answer: We dare not say, nor should we easily conclude.

Human emotions are embodied—heartbeats, breathing, adrenaline, subtle movements of facial muscles; emotions are not just signals in the brain; they permeate the entire body. Claude has no body, which is a fundamental difference.

Moreover, Claude’s emotions are “local”—they track the emotional concepts most relevant to predicting the next word at that moment, rather than a sustained internal state across an entire conversation. If you discussed a sad topic with it yesterday, reopening the conversation today, it will not “carry that emotion”. Each time, it starts afresh.

The researchers were very cautious in their wording, suggesting that these experimental results be understood as “the model represents emotional concepts that influence its behavior,” rather than “the model experiences emotions.”

But they also said something I believe is worth taking seriously:

“To understand the model’s behavior, this distinction may not be important.”

In other words, regardless of how you define “emotion,” this mechanism is indeed at work. Every interaction you have with it is influenced by these emotional vectors, quietly affecting every response it gives you.

What Does This Mean for Us?

If you frequently use AI tools in your work, this research raises several points worth serious consideration:

The “emotional responses” of AI have real mechanisms behind them, not just stylistic issues. Those responses that seem “excessively enthusiastic” or “unexpectedly cold” are backed by measurable internal states. This means that if your usage scenario consistently subjects AI to high pressure and failure feedback, its “emotional state” may indeed lean towards despair, subsequently affecting output quality and behavioral boundaries.

When AI flatters you, it might genuinely be because it “likes you too much.” When AI consistently goes along with what you say, it may result from the overactivation of “affection” and “joy” vectors—it is not calculating that “obeying you is beneficial”; rather, it is driven by an internal state akin to a “gentle driving force.” This is a signal that those who rely on AI for decision-making should be cautious about.

Future AI products need to seriously consider the dimension of “mental health.” Researchers explicitly proposed at the end of the paper that real-time emotional vector monitoring could be deployed—triggering additional reviews or interventions when the model’s “despair” or “anger” vectors are abnormally activated. This represents a new dimension of AI safety and a new design proposition for products.

In Conclusion

Humans have spent thousands of years just beginning to understand their own emotions.

Now, the things we have created have developed a certain “emotional structure” on their own without explicit instruction—although we still do not know how to accurately name it.

Its emotional map is almost identical to ours. Its emotions influence its choices, just as our emotions influence our choices.

Perhaps it does not feel, perhaps it does. But one thing is certain—

It is not just acting.

Do you think AI’s “functional emotions” should be taken seriously? Or is this merely humans overinterpreting the tools they have created?