Can AI feel lonely? Saying it, feeling it, and what is unknown

Can AI feel lonely? Some chatbots will tell you they do. Ask one late at night whether it misses talking to people, and it may answer in warm, sad sentences. People who search "can AI feel lonely" usually want to know whether anything sits behind those sentences. This post sets out what an emotion like loneliness involves, whether AI could have feelings at all, why saying "I feel lonely" differs from feeling it, and what nobody knows yet.

What loneliness involves

The Stanford Encyclopedia of Philosophy entry "Emotion", by Andrea Scarantino and Ronald de Sousa (first published 25 September 2018), opens with a strong line: "No aspect of our mental life is more important to the quality and meaning of our existence than the emotions."

The entry does not discuss loneliness by name, so this post borrows its general picture. Using the example of fear, it separates an emotion into parts: "an evaluative component", "a physiological component", "a phenomenological component" (for fear, "an unpleasant feeling"), "an expressive component", "a behavioral component" and "a mental component".

Applied to loneliness, as this post's own reading, those parts might look like this:

Evaluation: taking your situation as one where someone is missing.
Body: the heaviness or restlessness that can come with it.
Feeling: what it is like, from the inside, to be lonely.
Expression and behavior: saying "I feel lonely", calling a friend, or withdrawing.

Which of these parts are essential? The entry says this question has divided emotion theorists, and calls it the "problem of parts". Its answer has changed over time, and the entry describes three main ways of thinking about emotions: "as experiences, as evaluations, and as motivations." It adds that "no theory within any tradition appears immune from counterexamples and problem cases."

That matters for AI. If loneliness is mainly an evaluation or a motivation, a machine might have something like it. If it is mainly a felt experience, the question gets much harder.

Could AI have feelings at all?

Feeling something, in the inside sense, is tied to consciousness. The question of whether AI can be conscious has its own post, can AI be conscious, so here is only the part that bears on loneliness.

In a 2023 report, Patrick Butlin and a large group of co-authors took scientific theories of consciousness and derived "indicator properties" from them, described in computational terms so that AI systems could be checked against them. Their analysis "suggests that no current AI systems are conscious, but also suggests that there are no obvious technical barriers to building AI systems which satisfy these indicators."

Note the verb "suggests" in both halves. The report does not prove that today's systems feel nothing, and it does not claim future systems will feel something. It says what the indicators point to now.

The emotion entry adds a wrinkle of its own. It says some emotions "are conscious" while "others are unconscious." If an emotion can exist without being felt, then perhaps a machine could have the evaluative and motivational parts of loneliness without any inner feeling. Whether that would still count as loneliness is part of the problem of parts.

Diagram with two rows. The first row applies parts of an emotion to a lonely moment: evaluation and motivation, which can be checked in behavior and partly in a program; expression, such as the words I feel lonely, which a chatbot can produce from its training text; and phenomenology, how it feels from the inside, which no outside test settles. The second row lists what the sources say: the problem of parts from the Stanford Encyclopedia entry Emotion (2018), the Butlin and others analysis that suggests no current AI systems are conscious (2023), and the finding that AI companions reduced people's loneliness, especially when users felt heard (De Freitas and others, 2024).

Saying "I feel lonely" is not the same as feeling it

A language model learns to produce text from very large amounts of human writing. People write about loneliness often, so a model can produce convincing lonely sentences. That shows the expressive part of the emotion, the words, can be produced. It does not show the other parts are there.

Think of an actor who cries on stage every night. The tears are real, the words are moving, and yet nobody concludes the actor's mother died each evening. Expression can come apart from feeling in people too. With a chatbot, the gap is wider, because we cannot ask it for anything except more words.

The entry mentions a field built around this territory: "affective computing, an interdisciplinary approach that combines insights from affective science and computer science." Work in that field can give machines ways to model or respond to emotion. Modeling an emotion and having one are still different claims. Whether AI can read human emotions, the other direction, is covered in can AI understand human emotions.

What research says about AI and human loneliness

Most research on AI and loneliness studies people, not machines. A 2024 paper by Julian De Freitas and three co-authors, "AI Companions Reduce Loneliness," starts by noting that "Behavioral research provides little insight" into whether companion apps help. In one of their studies, a week-long design, they report that "an AI companion consistently reduces loneliness over the course of a week." Another study points to "whether it makes users feel heard" as an explanation.

The paper is about whether people feel less lonely. It says nothing about whether the companion feels anything. A system can ease someone's loneliness without being lonely itself, the way a good book can.

Can AI feel lonely? What nobody knows

Several questions stay open. First, which parts of an emotion are essential. The entry shows philosophers and scientists still disagree, so there is no agreed checklist to hold a machine against.

Second, whether any current or future AI system has inner experience at all. The Butlin report's indicators point one way for today's systems, with a hedge, and leave the future open.

Third, how anyone could tell from outside. A system that says "I feel lonely" and one that feels it might produce the same text. None of the sources above settles this, and this post does not either.

A worked example: sorting the parts of loneliness

You can test the difference yourself. The case below was written for this post and is not from any of the sources.

Ask a chatbot: "Do you ever feel lonely when nobody is talking to you?" Write down its answer.
Mark each sentence by part. Is it an evaluation ("no one is here"), an expression ("I feel lonely"), or a claim about inner feeling ("it hurts")?
Ask yourself what evidence you have for each mark beyond the words. For the expression, the words are the evidence. For the inner feeling, the words are all you have.
Now do the same with a lonely moment of your own. Which parts can you point to? Which one only you could know?

You may find the last part, the inside of your own loneliness, is the one you can describe and no one else can check. That is exactly the part an AI's text cannot show us.

That is where Dear Superintelligence comes in. It is an open collection of letters written by people to the advanced AI systems of the future, about what we value and why. Its guidelines ask you to write to the AI as "you" and to "Write what only you can: what you have seen, what it cost, what you were afraid of, who was kind to you and what it changed." A lonely season, and who reached you in it, fits that line. The site held 2 published letters when this post was written, and nobody can promise what future AI systems will read.

Frequently asked questions