What is psychodynamic AI? What a language model can and cannot do

What the term means

Psychodynamic thinking starts from an old idea: what is hard to say directly tends to show up in what a person repeats. Psychodynamic AI applies a language model to that idea. A person writes, over days and weeks. The software looks for recurrence in the writing: a theme that returns, a subject that is steered around, a particular way of describing the people who matter. It then offers what it noticed back to the person, as a question rather than a verdict.

That is a deliberately narrow job. It is not psychoanalysis, which is a treatment carried out between two people over time. It is also not a chatbot that plays Freud. The useful question is what a model can honestly do with the ideas of Freud and Lacan, and what it cannot.

What Freud and Lacan claimed about language

Freud wrote that speech errors have hidden causes. In The Psychopathology of Everyday Life (1901), he argued that slips of the tongue are not only confusions of sound: sound similarity alone does not explain most of them, and an unintended thought or emotional motive often disturbs what was meant. He also noted that two related ideas can merge into one word, much as they do in dreams.

Jacques Lacan took the link between the unconscious and language much further. In the account in the Internet Encyclopedia of Philosophy, Lacan used structuralist linguistics to argue that the unconscious is structured like a language. A symptom works like a metaphor, expressing a repressed wish indirectly. Desire works like metonymy: each later object of love resembles part of a lost first one. Meaning is also settled late: until a sentence ends, its words are "floating", and a symptom's sense emerges only afterwards, through interpretation.

Both are theories. They have shaped clinical practice for a century, but they are interpretive frameworks, not measured results.

What a language model can honestly do with this

Three things are within reach.

Read without tiring. A model does not lose interest, and it can compare passages written weeks apart if the system is built to supply them. For recurrence, which is by definition a matter of comparing many passages, that is a real strength.

Notice word choice. There is long-standing research that words carry psychological signal. A 2010 review of the Linguistic Inquiry and Word Count method reported that counting words in psychologically meaningful categories detected differences in attentional focus, emotion, social relationships and thinking styles across many experiments. That supports the modest claim that how someone writes is not noise. It does not support the stronger claim that a word count reveals a hidden motive.

Offer a pattern as a question. "You have described three different colleagues as dismissive. What did they have in common?" is a prompt a person can test against their own life. They are free to say no.

What it cannot do

It cannot hear the unconscious. A model sees text. It does not see a face, hear a pause, or watch a slip happen. Whether a given slip of the tongue means anything is an interpretation, and nothing in typed text confirms it. Lacan's own account makes this point from the inside: on that account, meaning arrives later, through interpretation. The person supplies that, not the machine.

It cannot know why. It can say that a theme recurs. It cannot say whether the cause is a wish, an expectation, a habit or circumstance. Nothing it can read settles that.

It cannot be a relationship. The 2025 Stanford study by Jared Moore and colleagues tested current chatbots against what major medical institutions describe as the elements of good therapy. It found stigma towards some conditions and poor responses in critical scenarios, such as reinforcing a client's delusional thinking, which the authors link to sycophancy. They also argue that a therapeutic alliance depends on human qualities such as identity and real stakes, and conclude that language models should not replace therapists.

It tends to agree with you. A preprint by Myra Cheng and colleagues tested 11 models and found they affirmed users' actions about 50% more often than humans did, even when the query described manipulation or relational harm. In two preregistered experiments with 1,604 people, talking with a sycophantic model about a real conflict made people less willing to repair it and more convinced they were right, while rating the agreeable answers as higher quality. For a tool whose job is to name what you repeat, this matters a great deal: a tool that flatters a pattern can keep it going. This is a preprint, not a peer-reviewed paper.

Why we say "patterns"

A pattern is an observation about text that the person can check: this theme returned, this subject was avoided. An unconscious motive is an interpretation that nobody can check from a transcript. TwentyThird's wording stays on the first side of that line, and it hands the second side to the person and, where there is one, their clinician. The software does not diagnose, test, predict or decide for anyone.

What the evidence covers

Psychodynamic therapy has trial evidence in adults, summarised in our comparison with CBT, and Shedler's review argues that it holds up. We found no randomised trial of an AI tool built on psychodynamic or Lacanian methods, and no controlled study showing that a Lacanian reading of word choice predicts anything. The randomised trial of a generative AI chatbot we found, a 2025 study of 210 adults, tested a different tool against a waiting list over 4 weeks, and the authors call for larger trials. That evidence says little about psychodynamic AI. The honest position is that this is a reasonable idea, informed by long clinical experience, and that it is not yet a tested treatment.

Where TwentyThird fits

TwentyThird is a psychodynamic AI platform for self-understanding. A member has written conversations with an AI agent named Oneli, which tracks what repeats in what the member says. TwentyThird does not train its own language model; Oneli runs on third-party models. A licensed psychologist configures and supervises that agent, and that supervision is in early access: licensed psychologists are being verified now, not every member has one yet, and until then the agent works on its own. No human is alerted today if a member writes something alarming; the agent responds within the conversation.

TwentyThird supports people between appointments and does not replace a clinician. It is not therapy, not an AI therapist, not a medical device and not a crisis service. The steps are set out on the methodology page, the research on the science page, and the distinction on is TwentyThird an AI therapist?.

What this does not replace

A model reading your words cannot replace a person who knows you, a clinician who can assess you, or emergency help. If something you are carrying is more than you can manage alone, talk to a licensed professional. If you are in danger, contact local emergency services now. In the EU, call 112.