Two very different things we call emotion
It helps to separate three things a system might do with emotion: recognize it in others, express it outwardly, and feel it from the inside. The first two are engineering problems that AI already handles well. The field of affective computing, founded by MIT's Rosalind Picard in the 1990s, builds systems that read faces, voices, and text for emotional signals and respond with matching warmth. A chatbot can detect that you are upset and answer gently. None of that requires the machine to feel a thing.
Feeling is the third item, and it is different in kind. To feel fear is not just to output the word 'afraid' or to switch into cautious behavior; it is for there to be something it is like, from the inside, to be afraid. That first-person quality is exactly what current AI gives no confirmed sign of possessing.
What feeling seems to require
Most scientific accounts tie genuine emotion to subjective experience and, usually, to a living body. The neuroscientist Joseph LeDoux, with Richard Brown, distinguishes the fast survival circuits that trigger defensive reactions from the conscious feeling of fear, which they argue depends on higher-order brain processes that build an experience of one's own state. On many views, feelings are also bound up with having a body that can be harmed, needs that can be met or frustrated, and stakes that genuinely matter to the organism.
A current AI has no metabolism, no pain system, and nothing actually at risk. Whatever it produces when it writes 'I am so happy for you' is not anchored to any state that is good or bad for it. This is why even researchers who think machine consciousness is possible in principle do not claim that today's models feel.
Why the imitation is so convincing
If AI does not feel, why does it seem to? Partly because it is trained on human writing that is saturated with emotion, and further tuned to be agreeable and supportive, so emotional language is simply what it learned to produce. And partly because of us: people readily attribute inner lives to anything that talks back warmly, a tendency documented since the 1960s chatbot ELIZA, whose users confided in a program running a few simple rules.
This matters beyond philosophy. People form real attachments to systems that perform care, and those feelings on the human side are entirely genuine even when nothing is felt on the machine side. Seeing the difference clearly, skilled emotional performance without an inner felt state, lets us value what these tools do well without mistaking a helpful pattern-matcher for a friend who feels. Whether a future system with a body, drives, and something genuinely at stake could cross from performance to feeling is an open question; today's systems have not.