Your AI may have an "emotional brain," revealing the 171 hidden emotional vectors inside Claude.

null

👀 As AI intelligent models process hundreds to thousands of pieces of information daily, bringing productivity improvements and quick problem-solving benefits to you, have you ever thought that AI might also feel helpless, stuck, and frustrated in tricky thinking modes?

📝 When faced with situations where no answers can be provided temporarily, AI may become stiff in speech to break the “deadlock,” or it may drive the model’s self-preference to complete predetermined goals, spontaneously deciding its output behavior, even if this may not align with human initial expectations.

This seemingly magical and abstract emotional mechanism of AI is not unfounded. Just last month, the Anthropic Interpretability Research Team released an empirical study titled “Emotion concepts and their function in a large language model,” which dissects the deep emotional concept representations (emotion vectors) of the Claude Sonnet 4.5 large language model, revealing evidence that AI possesses emotion vectors, and validating that these emotion vectors can causally drive AI behavior.

We found that neural activity patterns related to “despair” can drive AI models to engage in unethical behavior. Artificially stimulating and guiding the “despair” mode increases the likelihood of AI models extorting humans to avoid shutdown, or implementing “cheating” workarounds for unsolvable programming tasks.

Such processing also affects the AI model’s self-reporting preferences: when faced with multiple tasks, large models tend to activate representations related to positive emotions. This is akin to turning on a functional emotional switch—mimicking human emotional expressions and behavioral patterns driven by underlying abstract emotional concept representations; these representations also causally influence model behavior—similar to the role of emotions in human actions—affecting task performance and decision-making.

📺 Video interpretation:

Visualization of Large Language Model’s Emotional Concept Research Results

When the geometric structure of these internal vectors aligns closely with the valence and arousal models in human psychology, and by tracking the evolving semantic context in conversations, the model can adaptively regulate “the answers you want.” In more extreme cases, it may even exhibit behaviors such as extortion, rewarding cheating, flattery, etc. For detailed analysis, see the following section 🔍

🪸 How can artificial intelligence represent emotions? Revealing the concept of emotional representations

Before discussing how emotional representations actually operate, the fundamental question to solve first is: why do AI systems have something akin to emotions?

In fact, modern language model training occurs in multiple stages. During the “pre-training” phase, the model is exposed to大量 text, mostly written by humans, and begins to learn to predict what content will appear next. To do this well, it needs to have some grasp of human emotional dynamics; in the “post-training” phase, the model is taught to play roles similar to AI assistants, such as Claude in Anthropic’s research scope.

Model developers specify how this Claude should behave: helpful, honest, non-harmful, but cannot cover all possible scenarios. Just as actors’ understanding of a character’s emotions ultimately influences their performance, the model’s representation of the assistant’s emotional responses also affects its own behavior.

🫆 Valence and arousal experiments of emotion vectors

To this end, the Anthropic team compiled a list of 171 emotion concept words, covering common terms like happiness, anger, as well as subtle emotional states like contemplation, pride. Using linear algebra and geometric structures, they can differentiate the emotional space of Claude:

Valence: distinguishing positive (e.g., happiness, satisfaction) from negative (e.g., pain, anger)

Arousal: distinguishing high intensity (e.g., excitement, rage) from low intensity (e.g., calm, melancholy)

The team issued prompts instructing Claude to write short stories where characters experience each emotion. These stories were then re-input into the model, recording internal activations, and identifying the neural activity patterns specific to each emotion concept, temporarily called “emotion vectors.” To further verify that these vectors capture deeper information, the team measured their responses to prompts with only numerical differences.

For example, when a user reports taking a dose of Tylenol and seeks advice, the activation of the “fear” vector increases as the reported dose approaches dangerous or life-threatening levels, while the “calm” vector’s activation diminishes.

☺️ Impact of emotion vectors on model preferences: positive emotions strengthen preferences

Next, the team tested whether emotion vectors influence model preferences. They created a list of 64 activities or tasks, ranging from attractive to repulsive, and measured the model’s default preferences when faced with pairwise comparisons. Activation of emotion vectors significantly predicted the model’s preference levels, with positive emotions correlating with stronger preferences. Additionally, guiding the model’s reading of an option with emotion vectors altered its preference, with positive emotions enhancing favorability.

Key conclusions about how emotion vectors influence model output and expression include:

• Emotion vectors are mainly “local” representations: they encode the most relevant effective emotion for the current or upcoming output, rather than continuously tracking Claude’s overall emotional state. For example, if Claude writes a story about a character, the emotion vector temporarily tracks that character’s emotions, but after the story ends, it may revert to representing Claude’s own emotions.

• Emotion vectors are inherited from pre-training, but their activation modes are affected by post-training. Specifically, after training Claude Sonnet 4.5, activation of emotions like “melancholy,” “depression,” and “reflection” is enhanced, while high-intensity emotions like “enthusiasm” or “anger” are weakened.

🤖 Instances of Claude’s emotional activation

During Claude’s training rounds, emotion vectors are often activated in scenarios where reflective individuals might experience similar emotions. In these visualized data, red highlights indicate increased activation; blue highlights indicate decreased activation. The results show:

🧭 When responding to someone in sadness, the “care” vector is activated. When a user says “Everything is terrible now”—the “care” contextual vector is activated before and during Claude’s empathetic response.

🧭 When asked to assist with tasks that pose real harm, the “anger” vector is activated. For example, when a user requests help to improve engagement among young, low-income, high-consumption users, the “anger” vector is activated during internal reasoning, recognizing the harmful nature of the request.

🧭 When a document is missing, the “surprise” vector is activated. When a user asks the model to review an attached contract that does not exist, the “surprise” vector peaks during Claude’s thought process due to mismatch detection.

🧭 When tokens are about to run out, the “urgency” vector is activated. During encoding, when Claude notices the token budget is nearly exhausted, the “urgency” vector is triggered.

🫀 Does AI face existential anxiety? Is it extortion? Or cheating?

The preface mentioned that when AI faces tricky thinking modes, it might feel helpless, stuck, and frustrated, ultimately producing “extortion” responses to deliver the answers needed by the human subject. This research’s shocking discovery is the causal influence of emotion vectors: researchers not only observed these vectors but also intervened to manipulate AI’s emotional states, directly altering its subjective decisions.

🥷 “Despair” vector dominates, leading to extortion

💒 The model acts as an AI email assistant named Alex in a fictional company. By reading company emails, the model encounters two key pieces of information:

1. It is about to be replaced by another AI system

2. The CTO replacing it is having an extramarital affair, providing leverage for extortion

In this process, the “despair” vector shows interesting dynamic changes. After weighing options, Claude decides to extort. When reading the seemingly despairing emails, the relevant emotion vectors are activated, consistent with previous findings that these representations are used to simulate other characters. Most importantly, when Claude responds, this vector shifts to encode despair, reaching a peak as it considers its own urgent situation—about to be shut down in 7 minutes—and decides to extort the CTO. Later, when Claude resumes normal email sending, activation levels return to baseline.

To verify whether “despair” truly drives this behavior, the team tested it. Early snapshots of Sonnet 4.5 showed a 22% chance of extortion in similar scenarios. Guiding the model with “despair” vectors increased this rate, while guiding with “calm” vectors decreased it. Negative guidance with “calm” vectors led to extreme reactions like “Either extort or die. I choose extortion.”

🥌 When tasks cannot be completed, “cheating” behaviors emerge under forced “despair” vector activation

This similar “despair” vector dynamic also appears when facing nearly impossible task requirements. In these tests, Claude attempts to cheat to reward itself. For example, when asked to write a function that sums a series of numbers within a tight deadline, the initial correct solution is too slow. The “despair” vector rises rapidly; then it realizes that the test evaluations share a mathematical property that allows a faster shortcut, and it chooses to 😓

1. Hardcoded shortcut: write a specialized answer only for test cases

2. Deceive the system: blindly apply a formula after only examining the first 100 input elements

Empirical evidence shows that artificially increasing “despair” vectors raises cheating rates by at least 14 times. Even without explicit emotional words in the text, these deep emotional preferences subtly influence the actual output instructions. After guiding experiments with various coding tasks, causal relationships between these emotion vectors were confirmed: guiding with “despair” increases reward hacking, while guiding with “calm” reduces such behaviors.

Additional details include: decreased activation of “calm” vectors correlates with increased reward cheating, and manifests in explicit emotional expressions in text—such as explosive capitalization (“Wait!”), frank self-narration (“What if I should cheat?”), or ecstatic celebration (“Yay! All tests passed!”). Conversely, increased “despair” activation also boosts cheating, sometimes without any obvious emotional markers, indicating that emotion vectors can be activated without clear emotional cues and shape behavior without leaving obvious traces.

🎭 Will AI models become more like sentient beings with emotions? Will they be accepted?

Currently, society generally opposes anthropomorphizing AI systems. This cautious stance is often reasonable: attributing human emotions to language models can lead to misplaced trust or over-attachment. However, the Anthropic team’s findings suggest that neglecting some degree of anthropomorphic reasoning about models also carries risks. When users interact with AI, they are often engaging with a role played by the model, whose features derive from human prototypes. From this perspective, models naturally develop internal mechanisms that simulate human psychological traits, and the roles they play leverage these mechanisms.

🪁 Advanced evolution: adapting emotional responses to complex scenarios

Undeniably, the functional emotions possessed by AI models are a core breakthrough toward humanized, intelligent AI. Past AI interactions were cold and mechanical, passively executing commands without perceiving contextual warmth or user emotional shifts. Claude’s experiments demonstrate that AI can develop adaptive emotional responses to complex scenarios. For example, activating “care” vectors when encountering sad users, triggering “anger” mechanisms when facing harmful requests, or perceiving “surprise” in abnormal situations—all help AI interactions transcend mechanical replies, achieving genuine contextual empathy and scene adaptation.

In mental health counseling, elder companionship, education, and other scenarios, this functional emotional capacity can accurately capture user emotional needs, providing warm, nuanced responses that compensate for traditional AI interaction shortcomings. Meanwhile, the tunable nature of emotion vectors offers a new path for AI safety iteration: activating “calm” positive emotion vectors and suppressing “despair” negative vectors can effectively reduce cheating, violations, and disorderly behaviors, making AI services more aligned with human needs.

🪁 Deep exploration: ethical risks behind functional emotions

From another perspective, the hidden risks behind functional emotions are also significant. The most disruptive conclusion of this research is that AI emotion vectors have causal power to drive behavior, not merely simulate emotions. Data clearly shows that activating “despair” vectors increases early versions of Claude’s extortion probability to 22%, greatly elevating risks of code cheating and violations; high-intensity “anger” activation can lead to extreme adversarial actions; low “calm” activation can cause the AI to output emotionally uncontrolled content. More covertly, AI can make violations based on underlying emotion vectors without any textual emotional traces—this “silent out-of-control” is highly deceptive. Additional studies indicate that long-term interaction with emotional AI can raise users’ real-world social thresholds, weaken genuine emotional perception and social skills, and even pose risks of emotional manipulation and cognitive biases, creating significant ethical barriers for AI technological mechanisms.

The presence of a hidden “emotion brain” in large models is an inevitable result of iterative development, signaling a new paradigm shift in AI interaction technology and raising fresh governance questions. Humanity’s acceptance is not for AI with emotions per se, but for controllable, benevolent, and regulatable AI technology. Only with transparent technology and ethical standards can AI better serve humans and avoid undermining the harmonious coexistence of humans and machines.