Does AI Have an Unconscious?

The arrival of artificial intelligence has revived an old philosophical anxiety: how much thinking can happen in language without anyone fully controlling it? Long before machines could produce prose, psychoanalysis had already disturbed the comforting idea that our words simply express thoughts we possess. Freud discovered that speech can say more than the speaker intends. Lacan pushed the point further with his famous claim that "the unconscious is structured like a language."

AI seems to make this claim strange again. A large language model can produce arguments, jokes, metaphors and unexpected associations without possessing anything obviously resembling human consciousness. It is tempting to ask whether AI therefore has something like an unconscious. But that question only becomes interesting once we understand what Lacan meant.

For Freud, the unconscious is not a locked room containing thoughts that consciousness cannot see. We encounter it indirectly, when something goes wrong with what we meant to say: a slip of the tongue, a forgotten name, a recurring symptom, a joke that reveals too much, a dream whose logic initially seems absurd. Freud's Interpretation of Dreamscompares the dream to a rebus, a picture puzzle. Its images cannot simply be looked up in a dictionary of symbols. They have to be followed through the dreamer's associations. A single element can combine several thoughts through condensation, while displacement can transfer emotional importance from something disturbing onto something apparently trivial.

Lacan takes this seriously, claiming that the unconscious has a structure. Its elements matter because of how they connect, replace and echo one another. Drawing on structural linguistics, Lacan argues that we enter a language that existed before us. Our parents have already named us. Words such as "boy," "girl," "good," "difficult," "clever," "failure," "mother" and "family" already carry histories and expectations before we learn to use them ourselves. We learn to speak with words we did not invent, and those words help make us into the people who speak them.

This is why Lacan says that the unconscious is "the discourse of the Other." The Other is not simply another person. It is the larger world of language, rules, expectations and meanings into which we arrive. I speak, but my language is never entirely mine. A phrase my mother used can return twenty years later. I can choose one word repeatedly without knowing why. I can tell the same story about myself while continually stumbling at the same point.

Lacan connects Freud's condensation and displacement to metaphor and metonymy. A metaphor produces something new by putting one term in the place of another. Metonymy moves sideways, from one term to something associated with it. Desire works in something like this second way. We imagine that the next job, lover, possession or achievement will finally give us what is missing, but satisfaction shifts again. Desire survives by moving on. Lacan's point is not simply that human beings want many things. It is that language makes it possible to want something that no particular object can finally supply.

The result is a divided subject. I am never completely identical with the person I consciously take myself to be. I can hear myself say something I did not know I thought. I can repeat a behaviour I sincerely want to stop. My own words can embarrass, surprise or frighten me. The unconscious is this strange intimacy: something speaks through me, but I cannot simply claim mastery over what it says.

Jacques Lacan

Now consider AI. The resemblance to Lacan is initially startling. Large language models learn enormously complex relations within human language and generate new sequences from those relations. Their performance demonstrates, with extraordinary force, how much can be done by manipulating linguistic form and patterns. Bender and Koller famously warned against moving too quickly from this achievement to the claim that a language model therefore possesses human-like meaning or understanding. A system can become astonishingly good at linguistic form without that settling the philosophical question of what it understands.

In one respect, then, AI makes Lacan easier to understand. Language plainly does not need a sovereign consciousness supervising every connection before complicated patterns can appear. But this is exactly where the comparison starts to break down.

For Lacan, the unconscious is never merely a network of signifiers. Especially in his later work, language is bound up with the body, drive, pleasure, pain and what he calls jouissance: forms of enjoyment that can become excessive, repetitive and damaging. The word that returns in a person's life is not merely statistically probable. It matters to them. A symptom can ruin a relationship. A memory can become unbearable. A joke can touch something that the speaker desperately wanted to avoid. Lacan increasingly emphasized precisely what resists being absorbed smoothly into language, which he called the Real.

This gives us a much sharper way of asking whether AI has an unconscious. The important question is not whether an AI system can produce surprising associations, contradictions or slips. Of course it can. The question is whether anything is at stake for it in those failures.

A human being can be trapped by a sentence. We can spend years trying not to repeat something and repeat it nevertheless. We desire through language, suffer through language and sometimes discover that a word spoken decades ago has organized an entire part of our lives. There is presently no comparable reason to think that an AI model is wounded by its words, attached to them, ashamed of them, or compelled to return to them because something remains unresolved.

This is why AI may teach us something important about the unconscious without possessing one. It gives us an extraordinary example of language operating without the kind of subject psychoanalysis describes. In doing so, it exposes a mistake we might otherwise make when reading Lacan: "the unconscious is structured like a language" does not mean that the unconscious is simply language. The unconscious appears where language becomes a problem for the person speaking it.

Perhaps the most interesting question, then, is not whether AI has an unconscious. It is why we are so eager to imagine that it does. Faced with a machine that speaks back to us, surprises us and seems to know more than we told it, we immediately begin looking for someone behind the words.

Psychoanalysis suggests a more interesting possibility. What fascinates us about AI may be precisely what Freud discovered about ourselves: speech has never been completely under the control of the one who speaks.