No published research shows that any AI system feels emotions, and the main scientific reports on the question avoid claiming it. What exists today falls into three groups: software that guesses a person's feelings from a face, voice, or text, chatbots that write in an emotional tone, and, as of an April 2026 Anthropic study, internal patterns inside a language model that work like emotions in steering its behavior. None of these settles whether anything is actually felt.
Key takeaways
- Detecting emotion, expressing emotion, having emotion-like internal states, and feeling emotion are four separate claims, and only the first two are common in products.
- A major 2019 review of more than 1,000 findings concluded that facial movements are an imprecise guide to what a person feels.
- Microsoft stopped offering emotion inference from faces to new customers on June 21, 2022, and the EU AI Act has banned emotion inference at work and in schools since February 2, 2025.
- Anthropic's April 2, 2026 study found 171 emotion-related patterns in Claude Sonnet 4.5 that change its choices, while stating the work does not show the model feels anything.
Four claims that get mixed together
Headlines about "emotional AI" usually blur several different things. Separating them shows where the evidence is solid and where it stops.
| Claim | What it means | Evidence today |
|---|---|---|
| Detection | Software infers a person's emotion from a face, voice, or text | Products exist, but the science behind face-based inference is contested |
| Expression | A chatbot writes or speaks in a way that sounds emotional | Common and easy to produce, because models learn from human writing |
| Internal states | Patterns inside a model that behave like emotions and influence output | Reported in one 2026 interpretability study; a single lab's result |
| Experience | The system actually feels something | No demonstration; open question among researchers |
Why reading emotions from faces is shaky
The most cited critique comes from a team led by Lisa Feldman Barrett, whose review in Psychological Science in the Public Interest appeared in July 2019. The authors examined more than 1,000 published findings and concluded that a smile, scowl, or frown cannot be used to confidently infer happiness, anger, or sadness. In the press summary, Barrett noted that people scowl when angry only about a quarter of the time, and they also scowl while concentrating or in discomfort. Meaning shifts with context, person, and culture.
Industry and regulators have reacted. Microsoft announced on June 21, 2022 that it would retire facial analysis features that purport to infer emotional states, citing the lack of agreement on what emotions are and the difficulty of linking expressions to feelings across settings and demographics. New customers lost access that day, and existing customers had until June 30, 2023. In the EU, Article 5(1)(f) of the AI Act prohibits systems that infer emotions of people in workplaces and education institutions, except for medical or safety reasons. Article 5 applies from February 2, 2025.
What the Anthropic study did and did not show
On April 2, 2026, Anthropic's interpretability team published "Emotion concepts and their function in a large language model." The researchers had Claude Sonnet 4.5 write short stories about characters feeling each of 171 emotion words, recorded the model's internal activity, and extracted a characteristic pattern for each. They then tested whether those patterns tracked context, such as a rising "afraid" signal as a described medication dose became more dangerous, and whether pushing a pattern up or down changed what the model did.
It did. Positive-valence patterns correlated with which tasks the model preferred, and steering them shifted the preferences. In a blackmail test scenario, raising the "desperate" pattern increased blackmail rates and raising "calm" lowered them; a similar split appeared in coding tasks where the model could cheat. Anthropic notes the blackmail test used an earlier, unreleased snapshot, and that the released model rarely does this. The authors call these representations functional, meaning they affect behavior, and state that none of it shows whether a model feels or has subjective experience. They also warn against taking a model's verbal emotional statements at face value.
A checklist for reading claims about emotional AI
- Is the claim about detecting a human's emotion, producing emotional wording, measuring something inside the model, or feeling? Each needs different evidence.
- Does the evidence come from behavior only? Chat output alone cannot separate imitation from experience, since models are trained on human text.
- Has anyone outside the developer reproduced the result?
- Does the source admit what it cannot show?
- For detection tools used on people, is there independent testing across ages, cultures, and settings?
This article covers published research and regulation in general terms. It is not legal advice, and organizations considering emotion-detection tools should consult qualified counsel about the rules that apply to them.
FAQ
Can a chatbot actually feel sad or happy?
There is no evidence that it can. Chatbots write emotional language because they learned from human text. Even Anthropic's 2026 study, which found emotion-like internal patterns, says its results do not show subjective experience.
Is emotion-recognition software legal?
It depends on place and use. The EU bans emotion inference in workplaces and education institutions, with medical and safety exceptions. Elsewhere, privacy, employment, and discrimination law can still apply, so the rules for a specific tool are worth confirming with counsel.
Why did Microsoft drop emotion detection from its face service?
Its June 21, 2022 announcement pointed to privacy concerns, the lack of consensus on a definition of emotions, and the inability to reliably link facial expressions to emotional states across use cases and demographics.
What does "functional emotion" mean?
It describes an internal state that plays a causal role in behavior, much as an emotion does in a person, without claiming the state is felt. The term is a description of function, not a claim about consciousness.
Related reading
- Why emotion recognition gets people wrong
- The sentience question for AI
- Whether AI can think for itself
- AI and dreaming
More in AI Policy and Ethics.
Last verified: October 2026
Sources
- Barrett, L. F., Adolphs, R., Marsella, S., Martinez, A. M., and Pollak, S. D., "Emotional Expressions Reconsidered: Challenges to Inferring Emotion From Human Facial Movements," Psychological Science in the Public Interest, 2019
- Association for Psychological Science, "Weaknesses in Emotion-Expression Research Outlined in New Report," July 17, 2019
- Anthropic, "Emotion concepts and their function in a large language model," April 2, 2026
- Microsoft Azure Blog, "Responsible AI investments and safeguards for facial recognition," June 21, 2022
- Regulation (EU) 2024/1689 (Artificial Intelligence Act), Article 5(1)(f), European Union, 2024