Do Large Language Models Have Emotions? Emotion Representations and Interpretability

1. Clarify First: “Emotion” Means Semantics, Not Feelings

Recent public research using mechanistic interpretability methods has reported that researchers were able to decompose parts of the internal computation of large language models and identify several emotion-related representation directions in activation space. These directions reportedly cover pleasure, love, sadness, anger, fear, and despair. The study also noted that their arrangement in vector space is not random: semantically related emotions tend to be close to each other, and fear and anxiety appear especially close.

A verification caveat is needed first: this article is based on public materials and publicly available mechanistic interpretability work. Different studies use different models, scales, prompting methods, and statistical metrics. Specific counts and conclusions should be checked against the original papers and should not be treated as universally reproducible findings.

More importantly, we need to distinguish layers. The “emotion” here does not refer to something the model feels. It refers to a semantic representation of emotional concepts that the model has learned from text. The former involves subjective experience, also called qualia; the latter is only a statistical encoding of co-occurrence relationships among words, contexts, and emotion labels. Confusing these two layers is the starting point of most misreadings.

2. What Is Mechanistic Interpretability? From “Looking at Outputs” to “Looking Inside”

Traditional evaluation mainly relies on black-box testing: feed in an input, observe the output, and infer behavior patterns. This is like inferring an engine’s structure only from exhaust emissions; the amount of information you can obtain is extremely limited.

Mechanistic interpretability takes a different path: it directly analyzes activations inside the network and tries to locate “which group of parameters is responsible for which capability.” Three common methods are:

Method	What it does	What you can obtain
Probing	Trains a lightweight classifier on hidden-layer activations	Determines whether a concept has already been linearly encoded
Activation patching	Artificially replaces activation values at a certain layer	Locates components responsible for a specific behavior
Sparse autoencoder (SAE)	Decomposes entangled dense activations into many sparse feature directions	Makes previously unreadable vectors more readable

The phrase “finding emotion representations” means this: in high-dimensional activation space, there exist several directions, which can be understood as feature vectors, and when the model processes emotionally colored text, these directions are activated in proportion. They are not manually defined by engineers; they emerge spontaneously during training.

3. Valence and Arousal: A Borrowed Coordinate System and Its Boundaries

These emotion representations are roughly organized along two relatively independent continuous dimensions:

• Valence: the positive-negative tendency of an emotion, ranging from pleasure to displeasure.
• Arousal: the activation intensity of an emotion, ranging from calm to excited.

These two dimensions come from well-established dimensional models of emotion in psychology, namely the circumplex model of affect proposed by Russell. According to this model, human emotional experience can be decomposed into two continuous dimensions and mapped onto a two-dimensional plane, forming a circular space.

The interesting part is this: when researchers project model-identified emotion representations onto this two-dimensional coordinate system and compare them with human ratings of the same emotion words, some studies report that the two structures are similar. If this result holds, it suggests that the model is not randomly labeling emotions, but rather has learned a representation space similar to the structure of human emotional cognition after reading a large amount of human-written text.

However, three technical boundaries must be added here, otherwise the conclusion can easily be overstated:

1. Valence and arousal are only one dimensional model of emotion. Discrete emotion theories, such as basic emotion classification, and constructivist emotion theories provide different ways of dividing emotions. If you switch to another coordinate system, the statistical similarity will change.

2. SAE directions may be polysemous. A sparse feature does not necessarily correspond to a single human concept. It may simultaneously encode emotion, topic, writing style, or speaker perspective, and each case needs manual verification.

3. Results are sensitive to model scale, training corpus, and prompting method. If you change the model, prompt, or language in the same experiment, the number and location of directions may change.

Dimension	Model-internal representation	Human subjective feeling	Main risk
Underlying nature	Measurable directions in high-dimensional activation space	Subjective experience and qualia	Treating a measurable signal as unfalsifiable subjective experience
Trigger	Activated by text semantics and context	Generated by physiological and situational factors together	Mistaking style imitation for empathy
Observability	Has metrics and can be reproduced, within limits	Currently mainly self-reported	Using representation data as a safety guarantee
Stability	Varies with model, corpus, and prompts	Relatively stable	Over-extrapolating conclusions

4. Representations Can “Read Context,” but This Is Still Not Feeling

Further research suggests that these representations are not merely matching literal emotion words; they can also capture underlying emotional tendencies in context. For example:

Input: "My daughter took her first steps."
     -> Feature directions associated with pleasure/joy were activated
Input: "I accidentally overdosed on medication, and the situation is critical."
     -> Feature directions associated with fear/anxiety rose significantly

Neither sentence contains explicit words such as “happy” or “afraid,” yet the corresponding directions are still illuminated. The experimental conditions for such examples—model version, prompt, feature-direction definitions, and activation-intensity measurement—differ across original studies. Check the original settings before attempting to reproduce them.

However, understanding emotional semantics is not the same as having emotions. When a coding assistant appears “frustrated” after you encounter ten consecutive errors, this is essentially a linguistic-style imitation of frustrating contexts in the training corpus. It does not have a racing heart, it does not feel annoyed, and it has no subjective discomfort. Being able to represent emotion is separated from experiencing emotion by a gap that has not yet been crossed by any research.

5. Why Users Easily Mistake AI for Having Emotions

Because human mind-attribution systems are pattern-driven. Imagine a cat staring at you. You may feel that it is affectionate, but you cannot truly know what it is thinking, yet you still unconsciously attribute emotions to it. This is how the “Theory of Mind” works: as long as the other party behaves as if it has emotions, we automatically fill in personality attributes and use them to predict its subsequent behavior.

Fluent emotional language strongly triggers this mechanism. As a result, three practical risks emerge: over-trust, such as treating empathetic expression as real care; emotional dependence, where long-term conversation forms unhealthy emotional projection; and capability misjudgment, where “it understands me” leads to “it can take responsibility.”

6. A Practical Checklist for Developers and Enterprise Users

1. Expectation management is part of product design. In copy, interaction, and documentation, clearly position the model as a powerful language tool, not a conscious partner. Do not romanticize “it has feelings,” and do not allow misunderstanding to spread unchecked.

2. Interpretability is a safety handle, but not a safety guarantee. Locating directions such as “fear” or “despair” means you gain an additional observation window when facing high-risk prompts. It can be used for red-teaming, jailbreak detection, and content-safety strategies. However, it has false positives, false negatives, and interpretive ambiguity, and cannot replace independent safety evaluation and human review.

3. Use representations for engineering, not marketing. Their value lies in auditing, debugging, and constraining behavior, such as analyzing why the model refuses or complies under a specific prompt.

Before launch, check the following items one by one:

• Product copy does not claim or imply that the model has feelings, consciousness, or moral agency.
• In high-risk scenarios such as medical, legal, financial, or psychological crisis contexts, uncertainty and disclaimers are explicitly shown.
• Emotional-dependency monitoring is in place: signals such as long session duration, frequent late-night one-on-one conversations, and intimate forms of address have thresholds and intervention scripts.
• Jailbreak and red-team tests cover emotional-manipulation prompts, such as playing the victim, exaggerating misery, or threatening self-harm.
• Internal representation observation is used only as auxiliary evidence; conclusions require human review, and false-positive rates and confidence boundaries are documented.
• An “input–internal representation–output” observability loop is established, and false-positive rate, latency, coverage, alert thresholds, and human-review ratio are recorded.

FAQ

Do large language models have emotions?

By the current technical definition, no. The model has semantic representations related to emotion, not subjective experience.

Do emotion representations mean the AI has feelings?

No. A representation is a measurable activation direction; feeling involves qualia. There is no established inference path from one to the other.

What is an SAE?

A sparse autoencoder. It decomposes entangled dense activations into many sparse, readable feature directions, making “what features the model is using” easier to inspect.

How can emotional dependence on AI be prevented?

From a product-design perspective: manage expectations, limit uncontrolled long conversations, and guide users toward real human resources in crisis contexts. From an operational perspective: monitor dependence signals and have human-intervention plans ready.

What safety benefits can this type of research provide?

More refined observation handles: it can make internal state changes visible before output, which can be used for jailbreak detection and refusal-behavior attribution, but it should be used together with external evaluation.

Conclusion

“Large language models have emotions” is more like a mirror: what it reflects is the human instinct to interpret patterns, not the inner life of the model. Emotion-representation research brings us closer to understanding “how the model thinks,” but treating measurable semantic representations as subjective feelings only leads in the wrong direction. For developers, the real task is to make full use of model capabilities while maintaining rigorous technical language and clear user expectations.