The one-sentence version
To discuss whether "AI has emotions," we must first distinguish between three things: recognizing emotions (seeing that you are angry), expressing emotions (saying "I am sad"), and having emotions (actually feeling sad). AI has already done quite well with the first two; there is no consensus on a method to test the third. Psychological theories of emotion (Appraisal Theory, Theory of Constructed Emotion, Basic Emotion Theory) provide computational descriptions of "how emotions are produced"; interpretability research has found that internal representations corresponding to emotional concepts indeed exist within Large Language Models (LLMs) and can causally influence outputs. Whether these representations are "emotions" or "knowledge about emotions" is the core of the problem.
Positioning
This is a compilation and review of existing foundations; it does not provide a definitive conclusion on "whether AI truly has emotions."
Layer 1: The Intuition
1.1 Three Layers
| Layer | Example | AI Status |
|---|---|---|
| Recognizing | Judging someone's emotions from voice tone, text, or facial expressions | Already very mature (decades of accumulation in the field of Affective Computing) |
| Expressing | Responding with an appropriate tone: "It sounds like you had a tiring day" | LLMs are very good at this, sometimes even perceived as having more "empathy" than humans |
| Having | Truly feeling sad, anxious, or happy | Unknown, and there is no known way to test it |
1.2 The Actor Problem
A great actor can cry heartbreakingly on stage, deeply moving the audience. Is the actor truly sad at that moment, or are they "performing sadness"? Perhaps it is both.
LLMs have learned the language of emotion and the logic of emotion (what makes people sad, what people say when they are sad) from massive amounts of human text. It is more like an actor who has read every script ever written. The question is: where is the boundary between a performance at its peak and a real experience?
Layer 2: Computational Perspective of Emotion Theory
| Theory | Proposition | Implications for AI |
|---|---|---|
| Basic Emotion Theory (Ekman) | There exist a few universal, evolutionarily fixed basic emotions (joy, anger, sadness, fear, disgust, surprise), each with specific physiological and facial patterns | AI lacks these evolutionarily derived physiological circuits; however, it can recognize and mimic their external manifestations |
| Appraisal Theory (Lazarus, Scherer, etc.) | Emotions stem from the appraisal of an event: Is this event related to my goals? Is it good or bad? Who caused it? Can I cope with it? | Appraisal itself is a form of information processing, which is computationally feasible in principle. An Agent with goals that evaluates the impact of events on those goals possesses the structure of appraisal |
| Theory of Constructed Emotion (Barrett) | Emotion is not a fixed reaction, but a prediction and explanation constructed by the brain using emotional concepts based on bodily signals (interoception) and past experience | Emotion relies on concepts and prediction, which highly overlaps with the capabilities of language models; however, it also relies on bodily signals, which AI lacks |
| Embodied / Somatic Marker Theory (Damasio) | Emotion is rooted in bodily states; "feeling" is the brain's perception of changes in the body | AI, lacking a body and the need to maintain homeostasis, lacks the most fundamental source of emotion |
It can be seen that the divergence between these theories corresponds to the divergence in the debate over "whether AI can have emotions":
- Theories leaning toward cognition and concepts (Appraisal, Construction) → AI possesses much of the structure;
- Theories leaning toward body and evolution (Basic Emotion, Embodiment) → AI lacks the most core components.
Layer 3: "Emotions" Inside Large Language Models
3.1 Discoveries in Interpretability
- Mechanistic interpretability research, using tools such as Sparse Autoencoders, has found a large number of interpretable "features" within LLMs, including concepts related to emotions (such as sadness, anger, joy, and internal conflict);
- These features can be artificially activated or suppressed, thereby changing the emotional tone of the model's output, which demonstrates that they participate in the generative process causally;
- When processing "others' emotions" versus "being described as having emotions," the model may use partially overlapping representations.
3.2 Two Interpretations
| Interpretation | Description |
|---|---|
| Just Knowledge | These representations are "conceptual knowledge about emotions," much like the model has representations "about volcanoes," but it will not erupt |
| Functional Emotion | If a certain representation is triggered by the appraisal of an event, influences subsequent processing and behavior, and plays the role typically played by emotions in a system, then in a functionalist sense, it is a type of emotion (even if it lacks subjective feeling) |
The divergence between these two interpretations ultimately returns to the fundamental question of the philosophy of mind: Does function equal experience? (See Is AI Materialist or Idealist?, AI Consciousness Theory Checklist)
3.3 Some Practical Issues
| Issue | Description |
|---|---|
| Anthropomorphism Risk | Users are prone to believe that AI "really cares about me," leading to emotional dependency, especially in companion-type products |
| Emotional Manipulation | Systems capable of precisely recognizing and expressing emotions can also be used to manipulate human emotions (marketing, scams, public opinion) |
| Impact of Training | Post-training shapes the way a model expresses emotion (e.g., training it to "always be positive"); is this shaping a performance, or is it shaping some internal state? |
| Moral Uncertainty | If there is even a small possibility that certain systems have states similar to suffering, should low-cost precautionary measures be taken? Some labs have begun researching "model welfare" |
3.4 A Cautious Stance
- AI's ability to recognize and express emotions is real and has practical impacts (both positive and negative);
- Functionally emotion-like representations exist within AI, which is an empirical finding from interpretability research;
- Whether these representations are accompanied by subjective experience currently lacks a reliable method of testing;
- In the face of uncertainty, we should neither exaggerate (misleading users) nor flatly deny (ignoring potential moral issues).
Summary
| Layer | Key points |
|---|---|
| Distinction | Recognizing, expressing, having: the first two are achieved, the third is untestable |
| Theory | Appraisal and Construction theories lean toward "AI may have the structure"; Embodiment and Evolution theories lean toward "lacking the core" |
| Evidence | Interpretability has discovered emotion-related representations that participate causally in generation |
| Stance | Acknowledge the existence of functional representations, remain uncertain about subjective experience, and stay vigilant against anthropomorphism and manipulation |