What AI Is Good At—and What It Is Bad At

Psychotherapy, Psychoanalysis, and the Limits of Conditional Consensus

People often talk to AI as if it were a therapist. At first, this seems easy to explain: AI is always available, does not judge visibly, does not interrupt, and can reflect one’s thoughts back in organized language. It resembles a friend, a journal, a confessional booth, and a search engine all at once.

But that explanation is too broad. Many of those qualities also belong to friendship. Friends listen, reassure, reflect, normalize, and help us think through problems. The more precise question is not why people talk to AI as if it were a friend, but why they talk to it as if it were a therapist.

The answer has to do with the kind of authority AI appears to have. AI does not simply say, “I personally think your reaction is understandable.” It seems to speak from the accumulated center of human discourse. It has absorbed a massive corpus of things people have written: advice, arguments, manuals, essays, therapy language, professional standards, popular wisdom, bureaucratic norms, academic discourse, and ordinary conversation. When asked a question, it produces something like a conditional consensus.

This is what AI is especially good at.

Given a situation and a set of constraints, AI can redraw the relevant bell curve and give an answer near the center of what would count as reasonable, competent, or defensible within that context. If I ask what web development framework is best for a certain kind of project, AI can synthesize the mainstream considerations: maintainability, hiring pool, documentation, ecosystem, complexity, scalability, hosting, long-term support, and developer experience. It is not merely averaging all opinions. It is filtering for a qualified discourse: what competent people would generally recommend under these conditions.

The same applies to financial advice, recipes, tax workflows, software architecture, medical triage, interpersonal communication, and countless other domains. The user supplies conditions, and the AI produces the most defensible answer within the narrowed field.

In this sense, AI is a “best practices” machine. It is excellent at normative synthesis under constraints.

This also explains why people use AI therapeutically. Much of what passes for AI therapy is not a deep encounter with the unconscious. It is norm-checking. The user asks, “Is it normal that I feel this way?” “Was I wrong to say this?” “Am I being unreasonable?” “Is this a trauma response?” “Should I set a boundary?” “How do I communicate this without sounding harsh?”

AI responds from the center of contemporary therapeutic discourse. It says, in effect: “Many people feel this way.” “Your reaction is understandable.” “It may help to set a boundary.” “You might be experiencing anxiety, burnout, resentment, grief, or attachment insecurity.” It translates distress into socially recognizable, psychologically acceptable language.

This can be useful. It can reduce shame. It can help someone articulate what they are feeling. It can produce scripts for difficult conversations. It can calm someone down. It can prevent them from doing something rash. It can turn a vague emotional knot into a set of manageable sentences.

But this is not neutral. AI is not merely helping the user think. It is often giving the user symbolic authorization. It is telling the user that their feeling, reaction, hesitation, resentment, or fear falls within the range of the normal.

That may be therapeutic in one sense, but it is also precisely where the danger lies. “Normal” is not the same as true. A feeling can be common and still evasive. A reaction can be understandable and still cowardly. A resentment can be psychologically explicable and still self-serving. A boundary can be healthy, or it can be a way to avoid responsibility. A request for validation can be a legitimate need for stabilization, or it can be an attempt to get permission to remain stuck.

This is where the distinction between psychotherapy and psychoanalysis becomes useful.

Here I am using “psychotherapy” not loosely, but as something distinct from psychoanalysis. In Lacanian terms, psychotherapy operates largely on the lower tier of the graph of desire. It works at the level of meaning, demand, identification, ego-coherence, adaptation, and the subject’s relation to the Other as a source of recognition. The patient brings distress, confusion, conflict, or dysfunction; therapy helps organize it into meaning. It restores some degree of coherence. It helps the subject function better within the existing symbolic coordinates.

AI is very well suited to this. It is good at taking a complaint and turning it into a meaningful narrative. It can normalize, reframe, interpret, advise, soothe, structure, and script. It can help a person say, “This is what is happening to me,” or, “This is how I should respond.”

By contrast, psychoanalysis is concerned with something else. It is not primarily interested in restoring coherence. It is interested in what disrupts coherence. It attends to the symptom, the slip, the joke, the dream, the repeated word, the strange detail, the contradiction, the excessive affect, the oddly charged object, the thing that does not fit.

In psychotherapy, the question is often: “What does this resemble, and what usually helps?”

In psychoanalysis, the question is often: “What in this speech resists being assimilated to what we already know?”

This difference exposes a general weakness of AI.

AI is trained to complete meaning. It paraphrases, summarizes, categorizes, normalizes, and produces coherent answers. That is why it feels intelligent. But psychoanalysis often requires the opposite: not completing meaning too quickly. It requires attention to the singular detail that may be destroyed by a general explanation.

Consider a simple example. A patient has anxiety around ice cream. A therapeutic or psychological explanation might go toward food, pleasure, indulgence, childhood comfort, guilt, body image, dairy, sweetness, or contamination. These are reasonable associations. They are plausible. They belong to the field of general meaning.

But suppose the analytic work eventually reveals that the phrase “ice cream” matters because it contains “I scream,” and “scream” is connected to the patient’s childhood trauma of her mother constantly screaming. The anxiety is not about ice cream as an object. It is organized by a private signifying chain: ice cream, I scream, scream, the mother’s screaming, the child’s terror.

This is not population-level pattern recognition. It is not a best practice. It is not the center of any bell curve. It is singular.

The important thing is not that many people associate ice cream with screaming. They do not. The important thing is that, for this subject, this contingent phonetic link became psychically charged. The symptom is organized through language, sound, memory, trauma, and personal history in a way that no general model could infer simply by knowing what ice cream usually means.

This is what psychoanalysis trains the analyst to hear.

Human analysts do not solve this problem by having more data. They do not possess a giant database of cases where “ice cream” means “scream.” That would be absurd and anti-analytic. Analysts are trained through their own analysis, supervision, theory, seminars, and clinical practice to listen differently. The training is not primarily informational. It is formative.

The analyst must learn to resist the ordinary impulse to understand too quickly. This is harder than it sounds. Most listeners want to help. They want to comfort, explain, advise, reassure, classify, or make sense. The analyst must learn to suspend those reflexes. They must learn to hear exact words, repetitions, slips, metaphors, jokes, silences, contradictions, and strange intensities. They must learn to let the patient’s speech remain unresolved long enough for something singular to appear.

This is almost the inverse of the default behavior of an LLM.

LLMs are coherence machines. They are trained to produce plausible continuations, useful summaries, and socially legible answers. They are rewarded for being helpful. But psychoanalytic listening often requires being unhelpful in the ordinary sense. It requires not giving the user the answer they are asking for. It requires not translating the singular into the general. It requires not laundering the symptom into acceptable therapeutic language.

This does not mean AI could never be trained to assist psychoanalytic work. It could, in principle, be trained to track signifiers over time. It could notice that a patient repeatedly uses the words “clean,” “stain,” “spoiled,” and “smell” around sex, food, and the mother. It could detect homophones, repetitions, unusual metaphors, and affective shifts. With audio and video, it could notice pauses, vocal strain, forced laughter, changes in tempo, gaze, posture, or agitation. With biometric data from a watch, it could correlate speech with sleep, heart rate, movement, and physiological arousal.

In some ways, AI could become more consistent than human clinicians at noticing patterns. It could say: “Your speech rate increases whenever you discuss your father.” “You laugh when describing humiliation.” “You ask for reassurance after every moment in which responsibility is implied.” “The word ‘scream’ appears in three contexts you treat as unrelated.”

That could be powerful.

But detection is not interpretation. Noticing a pattern is not the same as knowing what to do with it. A patient may laugh when discussing abandonment because of shame, irony, contempt, dissociation, seduction, awkwardness, or habit. The fact of the pattern is not yet its analytic meaning. What matters is how and when it appears, how the patient hears it when returned, what resistance it produces, and what it does within the transference.

This is where human psychoanalysts may still have an advantage. Not because they have more data, but because analysis is not just information processing. It is an encounter structured by speech, desire, opacity, timing, and the subject’s relation to the Other. The analyst is not merely a pattern detector. The analyst occupies a position. The patient’s demand is enacted with the analyst, not simply described to the analyst.

Still, the important point is broader than psychoanalysis. The same limitation appears in many fields.

AI is strong where the task is to produce the best answer inside an established frame. It is weaker where the task is to notice that the frame itself is the problem.

This is what we usually call critical thinking.

Imagine an LLM trained only on material available up to 1926. If asked whether women should stay home and take care of children, it would likely produce an answer consistent with the respectable discourse of its time. It might say women have a special moral role in the home. It might support education for women, but as refinement for motherhood. It might be moderate, nuanced, and compassionate by 1926 standards. But it would still preserve the premise that gender is the relevant organizing category.

A genuinely critical response would ask: why should gender determine access to work at all? Who benefits from this arrangement? What economic, legal, religious, and symbolic systems make it appear natural? Why is “nature” being invoked here? What contradictions are being hidden by the appeal to family, morality, or social order?

An LLM trained on 1926 discourse would not reliably produce that critique unless the counter-discourse were sufficiently present in its training data or supplied by the user. It would not spontaneously step outside the dominant frame because the dominant frame would define what counted as reasonable.

This shows why “quality” is not enough. The highest-quality sources of a period may be precisely those most invested in the dominant ideology. A respectable answer can be worse than a marginal answer if respectability itself is part of the problem.

This is the general weakness: AI tends to confuse the center of available discourse with rationality.

It can reason very well from premises. It can apply constraints. It can produce sophisticated arguments. It can even generate critiques if asked to do so. But it does not inherently distrust the historical and ideological formation of the premises themselves. It does not naturally ask, “Why is this the frame?” unless that kind of questioning is already represented, requested, or rewarded.

This weakness appears everywhere.

In design, AI can give best practices for usability, accessibility, onboarding, navigation, and conversion. But it may miss the politically sensitive term, the institutional anxiety, the stakeholder contradiction, or the semiotic detail that determines whether the design works in the real situation.

In literature, AI can identify themes, influences, genres, and historical context. But it may flatten the singular force of a sentence, a rhythm, a tonal instability, or a word whose significance cannot be paraphrased.

In law, AI can summarize doctrine and generate arguments. But the decisive issue may be a small procedural wrinkle, a judge’s disposition, a local norm, or a phrase in a contract that matters precisely because it does not fit the standard template.

In medicine, AI can help with common differential diagnoses. But the diagnostic breakthrough often comes from the one symptom that does not fit the obvious pattern.

In business strategy, AI can produce a solid, mainstream plan. But the real opportunity may depend on an idiosyncratic founder, an odd distribution channel, a local cultural condition, or a market that standard categories misread.

In personal advice, AI can produce the reasonable-person answer. But the important thing may be that the person keeps asking a reasonable question for unreasonable reasons.

In each case, AI’s strength and weakness are the same. It is good at subsuming the particular under a general category. It is weaker at protecting the particular from premature subsumption.

This does not make AI useless. Quite the opposite. It clarifies where AI is most useful.

AI is excellent when you want conditional consensus: “Given these constraints, what is the most defensible answer?” It is excellent when you want best practices, mainstream reasoning, procedural guidance, comparison of options, summarization, reframing, or a first-pass analysis. It is excellent when the problem belongs to an established field and the goal is to act competently within that field.

AI is dangerous when the established field is itself the thing that must be questioned. It is dangerous when the correct answer depends on the anomaly, the exception, the singular detail, the historically excluded perspective, or the thing that does not yet have a stable place in discourse.

The lesson is not that AI cannot think. The lesson is that AI’s default mode is norm-production. It gives the answer that becomes reasonable after the context has been defined. Therefore the most important human task is often defining, challenging, or refusing that context.

In therapeutic use, AI can tell someone whether their feeling is common. It can help them function. It can give language to distress. It may even be better than many human therapists at consistency, patience, and practical reframing.

But psychoanalysis shows what this misses. The deepest question is often not, “Is this normal?” but “Why does this subject need this question answered?” Not “What does this resemble?” but “What is singularly at stake here?” Not “How can this be made coherent?” but “What does coherence conceal?”

That is the broader distinction.

AI is good at the center of the bell curve after the bell curve has been redrawn by constraints.

It is bad at the thing that does not belong on the curve yet.