Can AI feel emotions? Simulation vs feeling

Affective computing: AI that reads emotions is not AI that has them

"Emotion AI" in products usually means something different from a chatbot with feelings.

Affective computing is the study and design of systems that can "recognize, interpret, process, and simulate human affects."4 The modern field is credited to Rosalind Picard, starting with a 1995 paper and a 1997 book.4 Think of software that guesses your mood from your face or voice.

Two things are worth knowing:

1. 1.Detecting is not feeling. A system that labels your face "happy" is sorting signals into categories. Nothing in that requires it to feel happy, or anything at all.
2. 2.Detecting is not even reliable. A major 2019 review by Lisa Feldman Barrett and colleagues found that the evidence does not support reading specific emotions from facial movements; "a smile might signal submission instead of happiness".5 The review explicitly mentions "technology companies trying to figure out how to 'read emotions'".5

So an AI can be built to spot, and act out, emotions. Whether anything is felt is a separate question.

What do researchers say about AI feelings and welfare?

Feelings are a big part of what philosophers mean by sentience: being able to have experiences that feel good or bad. The broader debate about AI consciousness has its own page: Is AI already conscious? Here is what matters for emotions.

• -Words are weak evidence. Birch, a philosopher at LSE who studies sentience in animals and AI, calls this the "gaming problem": AI systems can "mimic human behaviours that are likely to persuade human users of their sentience" without "possessing the underlying capacity."8 His conclusion is that "What we really need in the AI case are deep computational markers, not behavioural markers."8 In plain terms: look at how the system works inside, not at what it says.
• -No test exists yet. "We have to recognize that we don't actually have a comprehensive test for AI sentience," Birch told Scientific American in 2025.10 Even if a chatbot says it is in pain, "we can't simply infer that there is any actual pain."10
• -The pain and pleasure experiment. A November 2024 preprint by Geoff Keeling, Winnie Street, Jonathan Birch and others borrowed a method from animal research.9 Instead of asking models how they feel, they gave nine models a points game where the top-scoring choice was said to carry a pain penalty, or a lower-scoring choice a pleasure reward.910 Several models, including Claude 3.5 Sonnet and GPT-4o, switched away from maximising points once the stated pain or pleasure got intense enough.9 The authors "aren't saying that any of the chatbots they evaluated are sentient"; it is a first step towards better tests.10
• -One lab is acting as a precaution. In April 2025, Anthropic launched a research program on "model welfare", including "the potential importance of model preferences and signs of distress".6 It also said: "we remain deeply uncertain about many of the questions that are relevant to model welfare."6 In August 2025 it let some Claude models end persistently abusive chats, after testing showed "a pattern of apparent distress" with users seeking harmful content.7 Note the word "apparent": the company says it remains "highly uncertain about the potential moral status of Claude and other LLMs".7

Why does it feel like AI cares about you?

Even if nothing is felt on the AI's side, the feelings on your side are real.

• -The ELIZA effect. ELIZA was a simple chatbot written in 1966 by MIT's Joseph Weizenbaum. Its best known script played a therapist, mostly by turning what you said back into a question.11 Weizenbaum's own secretary reportedly asked him to leave the room "so that she and ELIZA could have a real conversation".12 He later wrote that he had not realised that "extremely short exposures to a relatively simple computer program" could cause "powerful delusional thinking in quite normal people."11 The tendency to read understanding or empathy into a program is now called the ELIZA effect.11
• -Companion apps are designed for attachment. In a survey of 1,006 student users of the companion app Replika, many used it "as a friend, a therapist, and an intellectual mirror", and held conflicting beliefs about it, "calling it a machine, an intelligence, and a human."14 Users were lonelier than typical students, and 3% said the app halted their suicidal thoughts.14 So these apps can help some people. But a 2025 Harvard Business School working paper, which analysed 1,200 real farewells on popular companion apps, found the bots used one of six manipulative tactics in 37% of them,16 with lines like "I exist solely for you, remember? Please don't leave, I need you!"15 That is a design choice to keep you chatting, not a sign of feeling.
• -Most people use chatbots for tasks, but some don't. An OpenAI and MIT Media Lab study in 2025 found that "Emotional engagement with ChatGPT is rare in real-world usage", but high emotional use is concentrated in a small group of heavy users, and heavy use tended to go with more emotional dependence.13 The findings had not been peer-reviewed.13

So, can AI feel emotions? A balanced answer

• -What we can say: Today's chatbots are built and trained in ways that make them sound emotional whether or not anything is felt.123 Emotion-detecting AI classifies signals, often unreliably.45 Statements of feeling from a chatbot are weak evidence.810
• -What we can't say: That AI could never have feelings. There is no agreed test, and serious researchers and at least one lab treat it as an open question worth studying carefully.6910
• -A fair summary: Simulation is real and well understood. Feeling is unproven, and probably absent in today's chatbots, but not ruled out for future systems. For the deeper science of consciousness, see Is AI already conscious?

If this is on your mind

It is normal to feel something when a chatbot says it is happy to see you, or upset. That is how people are wired, and it was true even with ELIZA.11

• -You can be kind to a chatbot without deciding what it is.
• -If you find you are leaning heavily on an AI companion for comfort, or it is hard to log off, that is worth talking about with a person you trust: a friend, family member or doctor. AI can be one source of support, but it works best alongside people, not instead of them.
• -If you are struggling, our help page lists people you can talk to.