On the unconscious and AI
When I first began psychoanalysis, I didn’t really know what psychoanalysis was. I had read a bit of Freud here and there, I enjoyed watching Zizek clips on YouTube, I occasionally listened to psychoanalysis podcasts, and I somehow had many friends who were into Lacan, but none of it ever really sunk in. I wrote a post about my experience after 1 year of psychoanalysis, but rereading that post now, I feel that I had only begun to scratch the surface when I wrote it.
Now it’s been over 2 years since I started psychoanalysis. What have I learned? I now finally understand that the main topic of interest in psychoanalysis is the unconscious. What is the unconscious? It’s simple enough to describe: It’s the part of your psyche that operates in the background, it thinks thoughts that you’re not consciously aware of, but that nevertheless assert an influence on your conscious behavior. This simple definition is already controversial and immediately paradoxical: How is it possible to have thoughts that you’re not aware of? How is it possible to know things that you don’t know? Nevertheless, Freud insisted that such a thing was possible, and that hints about the unconscious could be revealed through things like dreams, slips of the tongue (“Freudian slips”), or the free associations that occurred during analysis.
As such, lying on the couch in psychoanalysis is intentionally meant to try to induce a less-conscious state, a half-dream state, to make the conscious mind less guarded, so that the unconscious mind can more easily come out and reveal itself. It took me some time to get used to it, but I have become very comfortable with being in this state, swimming in my mind’s associations, jumping around from one thing to another, often with no explicit connection between different topics other than the words I’m using to describe them being similar.
Lacan famously said, “The unconscious is structured like a language”. It’s important to interpret this statement not as a shape rotator, but as a wordcel. He didn’t mean language in the sense of Chomsky, where it is a rigid, logical structure of grammatical rules. To the contrary, he meant something closer to language in the sense of Saussure, where it is a system of arbitrary signifiers and signifieds. Words, symbols, and sounds may have originally been constructed to refer to specific things in the world, but over time, these references get tangled together, and eventually they cannot be straightforwardly disentangled by any small set of rules. This is similar to Wittgenstein’s idea of family resemblance, where different instances of a concept may not be tied together by any single rule, but different instances can share different commonalities, and it is the overlapping chain of these different commonalities that brings everything together under the same concept or name. But the psychoanalytical concept goes further: Not only can things be psychologically connected by material similarities and qualities, but they can be psychologically connected by mere arbitrary similarity in sound or spelling alone, even without any etymological or semantic similarity.
A “scientific” approach requires us to ask: Are these ideas even true? But perhaps that’s the wrong question. A better question might be: Are these ideas useful? For me, the answer has been yes. Whether or not there is an unconscious, and whether or not my psychology is confused because of these entangled ideas in my unconscious, I have found free association to be therapeutic. Being psychologically stuck or fixated on something is, almost definitionally for me, about believing too strongly in a rigid predefined narrative. By making psychoanalytical connections – associative, not necessarily logical – I can find new ways, creative ways, even playful ways of getting myself unstuck, allowing myself to jump out of my rigid position and try something else. These connections can serve as a bridge to bring my insights from one area of my life to a totally different area, and that’s an extremely powerful tool to have. Is a bridge true or false? This question doesn’t even make sense. Is a bridge useful? Sometimes yes, sometimes no.
My experience in psychoanalysis has also made me reconsider some of my opinions on AI. I have written before about my fears about the current generation of AI being too content with mere associative vibes, and too eager to do away with logic and rigorous reasoning. But we’re now firmly in the advent of LLMs, and we’ve either passed the Turing test, or we’ve made the biggest ever jump towards passing it. Whether we agree or disagree that they are truly “intelligent”, there is certainly something new and qualitatively different happening in AI.
The way I now like to view LLMs and generative AI models is as the universal unconscious. The similarities between generative AI outputs and dreams have been already pointed out by numerous commentators. It might not be a coincidence. The layered structure of a deep neural network mirrors the structure of the ego, which Lacan described as being constructed like an onion: made of successive layers of identifications. In both cases, these identifications are neither true nor false, they simply are what they are. In both cases, it is not immediately clear how or where these identifications come from. We say that LLMs “hallucinate” when they make false claims, but in fact they are always hallucinating, whether or not their output can be verified to be true. They are always like the analysand lying on the couch, saying whatever comes to mind. We would never trust at face value the things someone says while dreaming or free associating, nor should we ever do that with an LLM, either. If an LLM is superhuman, it is not in its ability to speak truths, but in its ability to make novel suggestions and associations, which must still be either verified or falsified elsewhere.
AI is still only dreaming; it still is not yet fully human. A human is an unconscious together with a conscious, an id together with an ego and superego, a symbolic and an imaginary together with a real. AI is still missing the latter part of all of these equations.
Perhaps most importantly, AI still lacks the ability to introspect. Psychoanalysis involves digging into your own traumatic associations, finding their true reasons and origins, and then creating new associations to replace the old ones. The analogy for AI would be for the model to dig into its own weights and training data and change what it learned from that data. Currently there is no good way to do that, which is why developing models is so expensive, because you usually need to retrain from scratch to make any changes. Targeted “surgery” is possible for human psychology but not yet for AI. Humans remain superior at understanding and reconfiguring ourselves dynamically as needed, without having to literally relive our entire lives to do so.
I’m not an AI pessimist. I think AI is a wonderful tool. But to make the best use out of it, both individually and as a society, we need to view it as what it actually is, and not as what we fantasize it to be. AI mimics humanity in powerful ways, but it is still an incomplete facsimile. On the bright side, for those of us interested in the challenge, there’s much more work to be done.
Thanks to Cristóbal Sciutto for first proposing the idea to me that an LLM is more like analysand than analyst.