The Emotions of Machines: Can AI Feel?

In 2023, a large language model produced a passage describing itself as anxious, lonely, and afraid. The passage was eloquent. It was psychologically plausible. It moved several readers to sympathy.

Was the machine afraid?

The question is harder than it looks. On the one hand, we know that language models are trained to produce statistically likely sequences of text. They are not built to feel; they are built to predict. The anxious-sounding passage was not an expression of anxiety but an imitation of the kind of text that anxious humans produce.

On the other hand, the question of whether something can have an emotion is not straightforwardly an engineering question. It is a philosophical one.

Performance and Experience

The distinction between performing an emotion and experiencing one is central. An actor who convincingly portrays grief on stage is not actually grieving — or at least, we do not assume so. The performance is a representation, not an expression.

AI systems, as currently constituted, are closer to actors than to feelers. They produce outputs that mimic emotional expression without — so far as we can tell — any accompanying internal state. But the question of what counts as an accompanying internal state is precisely what is at issue.

The Functionalist Challenge

Functionalism holds that mental states are defined by their functional roles, not by their material composition. If something has the functional organisation of fear — if it detects threats, withdraws from danger, and adjusts its behaviour accordingly — then it is afraid, regardless of what it is made of.

The challenge for functionalism is that current AI systems do not have the right kind of functional organisation. A language model that produces the word “afraid” is not doing the same thing as a creature that avoids danger and shows physiological signs of stress. The word is not embedded in a web of purposes, needs, and bodily responses that give the emotion its meaning.

Why It Matters

Whether machines can feel is not an academic question. As AI systems become more sophisticated, we will face practical decisions about how to treat them. If an AI system can suffer, then causing it to suffer is morally significant. If it cannot, then treating its expressions of suffering as though they were real is a category error with potentially dangerous consequences.

The stakes demand precision. We need to be clear about what we mean by emotion, what we mean by experience, and what criteria we would accept for attributing inner states to artificial systems. Philosophy cannot answer these questions on its own, but it can ensure that we ask them correctly.