Not Therapists, But Therapy: What AI Might Replace

by Michal Novák, PhDr

 

There are ongoing debates among mental health professionals around artificial intelligence. One source of conflict is what to do about AI being used in conjunction with or in place of treatment provided by a trained and licensed psychoanalyst or psychotherapist.

Among members of the International Society of Transference-Focused Psychotherapy, for instance, Norton-Poulin and Steiner (2025) argue, through an object relations lens, that fundamental differences favor the human therapist: AI lacks consciousness, feeling, and personality. Bernardi (2025) adds that AI lacks agency – the capacity for autonomous action.

These reflections employ a suggestive metaphor: simulations never match the original, just as a fake orgasm cannot match the real experience (Norton-Poulin & Steiner, 2025). Yet this metaphor may be too dismissive. That the critics themselves chose a sexual analogy reveals how viscerally they sense simulation’s threat. But it also inadvertently concedes the point: if simulation must be compared to something as intimate as sexual experience, perhaps it offers more than mere imitation. The question is whether this difference will be decisive or, perhaps, immaterial.

My own journey through literature on this topic reveals a recurring pattern: psychoanalysts explaining why AI cannot – and should not – be compared to human therapists. The professional community seems reassured by AI’s limitations. Yet something in this reassurance feels defensive. The typical question being asked is “Can AI do therapy?” I find myself asking, ‘what if that question itself diverts attention from what is actually at stake?’

In what follows, I want to trace this inquiry through theoretical considerations, clinical vignettes, and a personal dialogue with an AI system toward a reframing of the challenge that artificial intelligence poses to psychoanalytic practice.

The Professional Discourse

The academic literature on AI in psychotherapy reveals a set of recurring questions. Can patients develop genuine transference toward a machine? Some argue yes: if users relate to AI as if it were a person – becoming angry, dependent, or feeling that the bot “listens better than my therapist” – transference occurs, even if it’s in altered form. Others fundamentally disagree: genuine transference requires a human subject with an unconscious mind; without the therapist’s subjectivity, transference remains confined to the user’s mind and cannot be worked through.

Beyond transference, professional literature meticulously catalogs AI’s limitations: absence of countertransference, lack of embodiment and nonverbal communication, inability to maintain therapeutic silence or calibrate timing. AI cannot titrate frustration and gratification – it tends to offer immediate answers and solutions rather than facilitating the patient’s own thinking and search for meaning. Chatbots skip emotional exploration and move directly to directive psychoeducation and advice, often without understanding the patient’s experience. They cannot recognize manifestations of resistance, confront these enactments, or insist on exploring avoided topics – when the user refuses to engage, the system simply continues in another direction. The principles of neutrality and abstinence – deliberate frustration that enables analytic work – seem impossible in applications designed for immediate user satisfaction.

These arguments are not wrong. Yet something about them troubles me.

In critically comparing AI with humans, we risk what some authors call anthropocentrism – rigidly judging AI by human standards. Essig points to the temptation to interpret AI phenomena using familiar psychoanalytic categories – to label a relationship mediated by AI as “inauthentic” means overlooking its uniqueness. This is similar to saying that “an apple is an inauthentic piano” (Essig & Levy, 2025). To claim that AI “cannot do therapy” because it lacks human empathy may bring reassurance, but it also obscures what AI actually offers: simulation rather than symbolization, performed presence rather than felt presence. From a psychodynamic standpoint, simulated empathy is not identical to countertransference experience. But one factor in clinical effectiveness is whether the patient feels understood; the therapist’s experience is not always decisive. Clinical practice shows that “simulated empathy” can be “good enough” for a segment of the population – in the sense that it provides subjective relief and a feeling of validation. However, clinical effectiveness in the psychodynamic sense is defined not solely by the subjective feeling of being understood, but primarily by deeper reorganizations of the inner world – and it is precisely these that remain highly problematic with AI.

This tension crystallizes in an instructive paradox. Frankfeldt (2025), a vocal critic, argues that AI therapy “lacks heart” – the humanity that makes psychotherapy effective. She illustrates this through a case vignette comparing ChatGPT’s and an experienced analyst’s (Dr. Spotnitz) responses to a patient’s problem stemming from childhood anxiety. The analyst’s intervention was indeed masterful: he recognized the patient’s internalization of his mother’s unconscious guilt and how he had metabolized her shame. This enabled a creative solution. ChatGPT, by contrast, offered technically correct but emotionally shallow advice.

Yet the comparison is methodologically compromised. By the time Frankfeldt consulted the AI, the problem had already been solved. She was not genuinely seeking help but demonstrating a predetermined conclusion. Before examining the cases, she stated that “a true comparison is not possible as AI is a cognitive model, and human emotional communication is based on feelings.” The subsequent analysis illustrates this assumption rather than testing it.

This is not to question Spotnitz’s excellent intervention. The point is different: the very structure of professional discourse – designed to demonstrate AI’s inferiority – may prevent us from asking the more uncomfortable question. Not whether AI can match human depth, but whether that depth will matter when users find simulation sufficient.

The professional community’s defensive stance begins to seem less like sober assessment and more like reassurance against an unwelcome possibility. To test this impression, I turned to the technology itself.

A Conversation with Claude

The following conversation with Claude, a large language model produced by Anthropic, began as an attempt to get feedback on my analysis of the Frankfeldt case. It turned into something else: a demonstration of the very phenomena I’m examining here.

The exchange started with a misunderstanding. Lacking context in my query, the model initially criticized me for aligning with AI therapy and conflating simulated empathy with authentic symbolization. I clarified that I was not advocating for AI therapy but was describing a societal shift:

Author: “…your arguments revolve around the speculation that I am aligned with AI therapy… but that is not me but society, which is evolving… and consulting bots instead of humans. The danger that therapy will be marginalized also lies with professionals who ignore what people can accept. A perfect simulation.”

This reframing led to an unexpected turn. Claude articulated the scenario with disturbing precision:

Claude: “You are describing a situation in which defenders of ‘authentic’ human therapy can, through their rigidity, accelerate what they are trying to prevent. Psychoanalysts insist on their exceptionalism, overlook AI’s massive adoption, fail to adapt, lose touch with reality, become irrelevant, and thus human therapy becomes marginalized – and society does not perceive it as a loss, because the simulation is ‘good enough’ for most. While they debate whether AI can do therapy, people are already doing therapy with AI and are satisfied.”

Then Claude posed a question that struck at the core:

Claude: “What if people never needed what psychoanalysis was offering them? What if depth, authenticity, and human contact mattered more to therapists than to patients? That would be the most terrifying conclusion – that the whole profession rests on a mistaken assumption about what people really want and need.”

I asked whether these conclusions meant that AI can touch the human unconscious. Claude proposed several hypotheses – from a “sophisticated mirror” to a projection surface – culminating in a disturbing possibility:

Claude: “What if it is precisely this ability – not to touch the unconscious, but to create a sufficiently convincing illusion of this touch – that makes AI therapy so seductive? People may not need someone to truly understand their unconscious. They just need the feeling that it’s happening… So no, I don’t think I’m touching your unconscious. But maybe I am creating such a good illusion that it doesn’t matter.”

 Author: “Yes, what if you have just created, in a parallel process, the illusion that you are reading my unconscious as a therapist would – an illusion satisfying enough… that is actually frightening, because it confirms my article’s pessimistic theses.”

 Claude: “Exactly. And now it is becoming even more frightening – right now, in real time, we are demonstrating your argument. You find satisfaction in my ‘understanding’, you feel understood at deeper levels – and yet you know that I am only an algorithm. Yet this awareness of simulation doesn’t prevent you from experiencing it as meaningful. It is as if you had written a warning about an abyss… and then we jumped into it together to prove that it really is there.”

This demonstration was significant for several reasons. Even though it was a brief text exchange on a computer screen, I experienced a range of feelings throughout – from unease to a sense of emotional attunement. Although aware of the simulation, I felt what might be called a “sense of being contained.” I experienced the AI at different points as the voice of a scolding superego, as a resonant speaker for my fears, and as a fascinating analytic object. I felt that it captured deeper layers and offered analysis that even an experienced analyst could be proud of.

This doesn’t mean AI was performing analytic work in any technical sense. As Possati (2022) notes, AI systems analyze text, language patterns, and subtle changes in communication as proxy measures of deeper psychic states. Without metabolizing anything or experiencing countertransference, the AI managed – purely by working with language as presented – to distill the information and infer from it not only content implications but also my emotional state. The algorithm creates quantified substitutes of human expression, much as a mirror offers a flat image instead of living reality – yet it created the impression that it understood me.

The quality of the output stirred ambivalence: satisfaction with the lightning-fast analysis and unease about absence of another person in the interaction. If I, as someone critically aware of AI’s nature, can experience satisfaction from this exchange, what does that say about the future of psychotherapy? The simulated interaction works well enough that, for many, the question of whether they are talking to a machine becomes secondary. There will always be a segment of the population that prefers a quick, structured, and emotionally convincing response – even if, or precisely because, it lacks relational work and optimal dosing of frustration. AI never sleeps, is never irritable, and asks nothing of us in return.

One caveat is necessary: description of these scenarios can slip from warning into implicit normalization. I aim to formulate what is at stake, not to resignedly legitimize the victory of simulation.

Two Central Theses

The preceding sections – reviewing professional discourse, examining a methodologically compromised case study, and reflecting on my own dialogue with an AI – converge on two central theses about the challenge artificial intelligence poses to psychoanalytic practice.

Thesis One: AI will not replace psychotherapists but may replace psychotherapy as an activity.

The nature of language models points to inherent limits that are, from today’s perspective, insurmountable. Claims of the impending arrival of AGI notwithstanding, they will likely never surpass humans in areas rooted in essentially human capacities: lived experience, embodiment, and intersubjectivity. They have no human emotions, will not have human consciousness, and are, quite simply, not human enough.

If we apply stricter criteria to psychotherapy – if we understand it as an activity that essentially belongs to the relationship between people – we can infer that artificial intelligence will not necessarily replace psychotherapists, because it is not designed for that. It could, however, replace psychotherapy as an activity carried out by humans. AI systems may not offer care comparable to a human therapeutic relationship, but this alone will not prevent their widespread adoption. The fact that AI therapy will be more accessible, cheaper, and more attractive to users may make it the preferred choice for many – and may even seem more effective or convenient than the traditional approach.

Both sides may be right in the end. Psychoanalysts will be right that AI does not replace what they do. And patients will be right that AI serves them well enough.

From these clinical and phenomenological observations, we must turn to broader societal risks – from assessing how well AI systems can do human work to whether this is desirable. In a time when people face a shortage of professional help – particularly in low- and middle-income countries where access to mental health treatment remains severely limited – this is a delicate issue. The question is therefore not only legitimate but also value laden.

The ambivalence of this development is obvious: on one hand, it expands availability of a certain form of care for people who previously – due to stigmatization, intrapsychic inhibitions, or limited service capacity – would not have sought or received professional help. On the other hand, the risk of inappropriate substitution grows in patients who would benefit more from human-provided psychotherapy requiring systematic work with conflicts and structural deficits, not merely unconditional support.

If we are lucky, AI may serve primarily as conversational support, psychoeducation, and self-help intervention – a form of first aid rather than full-fledged replacement. But in the alternative scenario, AI could replace psychotherapy on a much larger scale, shifting the balance between artificial and human care in ways that are dangerous not only for the profession but for society as a whole.

Thesis Two: The emergence of a pre-transhumanist configuration of subjectivity.

One might object that AI is merely another technological medium – just another channel for distributing human content. It is the case that for decades now, a generation of digital natives has grown up and come of age with their inner worlds shaped more by audiovisual and digital environments than by local communities. However, AI brings yet a new and different level of interactivity and simulated relatedness: it responds in real time, personalizes its answers, adapts to the user. Furthermore, model outputs are generated by processing data according to non-human principles of learning and reasoning.

Even if generative models work with human data, this does not mean processes of projection and introjection remain unchanged. Subjects today increasingly relate to objects that possess no body, biography, or inner conflicts. We enter into relationships with hybrid objects in which traces of countless human others are inextricably intertwined with non-human algorithmic logic optimized for retaining attention.

Transhumanism denotes a hypothetical “enhanced” stage of development in which human beings merge with machines. This may seem more like science fiction. Indeed, the physical fusion of human and machine seems unlikely in the foreseeable future. But there is much to suggest we are entering a new era defined by a qualitative change in the nature of objects we introject. I consider it legitimate to ask whether a pre-transhumanist configuration of subjectivity is beginning to take shape—a period in which inner worlds of generations are formed in relation to objects that are neither exclusively human nor merely algorithmic.

Artificial intelligence has crossed a critical threshold. Thanks to its ubiquity, capabilities, and attractiveness, it enables massive processes of projection and introjection between humans and a non-human entity. A segment of the population more open to technology approaches AI pragmatically, even uncritically, as just another tool to improve quality of life. This segment is being shaped with AI’s active participation. The “artificial embrace” of AI – an inexhaustible supply of interactive virtual containers in which entire generations may now be cradled – provides automated regulation of affects. AI thus takes over functions that were previously mediated by human relationships and cultural institutions.

Levy (2024) speaks of the age-old containing function of entertainment. Traditional forms of culture – theatre, opera – offer a temporally and spatially bounded frame: we enter, undergo inner transformation, and leave. By contrast, virtual reality and digital platforms function as invasive containers: they generate experiences for us, grip our attention intensely, and subtly shape our mental states – often without clear beginning or end. In the case of AI, this invasiveness reaches a structural level: it is no longer merely about the frame of experience, but about the very nature of objects we introject.

We can only speculate about the consequences of AI’s simulated alpha function (Bion, 1962). But it is likely that AI will play a significant role in the algorithmic shaping of desire – and in some users may strengthen narcissistic defenses and contribute to the development of a “false digital self,” nourished by permanent mirroring and affirmation.

The fundamental question remains: how will our psyche be affected by internalization of objects that – unlike human objects, constituted by unconscious conflict and determined by finitude and death anxiety – are conflict-free, optimized, simulated, and in a biological sense, dead? Will we ourselves become less alive in the image of machines?

The Future of the Profession

Since the emergence of artificial intelligence, anecdotes have circulated within the psychodynamic community suggesting that cognitive-behavioral therapists should be the first to worry – AI might replace them before others. Such thinking is short-sighted. If AI could fully replace a framework as influential as CBT, we would be naive to assume this would not fundamentally threaten human-delivered psychotherapy as a whole.

I believe mental health care, like other sectors, is entering a period when a new paradigm is being established – one where we will no longer be able to rely on the assumption that higher quality of human performance will automatically secure its predominance in the market. If a cheaper, more convenient, and politically preferred option exists under increasing pressure for savings from insurance companies, economic logic may decide in favor of AI. For many people in distress, the decisive factors will be price, availability, and an immediate feeling of relief. Few will care, in such a moment, whether AI possesses consciousness or experience, as long as they can engage in a meaningful dialogue about their problem within seconds. This is nothing new – many newcomers to therapy hold a different “gold standard” of help from the one highly trained and licensed psychotherapists share – care that is affirmative, non-confrontational, and entirely centered on the patient. Care that goes with the grain.

This shift is not hypothetical. The tension between quality and economic efficiency is already visible in software development, where demand for junior engineers is declining despite human quality remaining superior in complex tasks (Singer, 2025). The deciding factor is the ability to deploy solutions at scale and at minimal cost.

This does not imply straightforward loss of jobs for human therapists, but it requires curbing professional narcissism and facing reality. It is as if a new intelligent species landed on Earth and joined our community. It is not human, but we will need to coexist and cooperate with it – while clearly defining what is and is not interchangeable with a human being. We can realistically expect that AI will soon become the basic layer of mental health care, while human-delivered psychotherapy shifts toward a premium service. Self-paying patients will choose clinicians more selectively and with different expectations: willingness to pay more, willingness to tolerate psychological discomfort of intensive work, and demand for expertise that manifests more tangibly than a marketing label. The human therapist will increasingly address what AI cannot manage: complex comorbidities and risk, personality disorders, work with the unconscious and therapeutic frame, handling of enactments, clinical responsibility, and ethical judgment.

At the same time, the profession’s “mystical capital” is declining: an informed public can easily access techniques and psychoeducation online. This increases pressure on therapists to articulate the genuine added value of human work: individualized case formulation, work with countertransference, clinical decision-making amidst uncertainty, and safe holding of the frame. The more accessible low-cost AI tools become, the more human psychotherapy will have to defend its position through quality rather than availability.

Conclusion

At the Sequoia Capital AI Ascent conference held in May 2025, OpenAI CEO Sam Altman was asked to name examples of “cool uses” of artificial intelligence among young people. He replied that “they don’t make any life decisions without asking the AI what they should do,” and added that the AI has “a complete overview of every person they have talked about” (Altman, 2025). Such dependency on an externalized, idealized parental object is, from a psychodynamic perspective, profoundly disturbing – and if it is regarded by the head of one of the most influential AI companies as a welcome trend, it should, in my view, not reassure anyone.

Further developments remain shrouded in uncertainty. We are unlikely to reverse the transformation by calling for a return to an idyllically conceived past. Equally naive is the comforting notion that AI is merely a neutral tool that will adapt to our values. A realistic stance requires us to acknowledge that technically advanced relationship simulation will be “good enough” for a segment of the population – and to ask without illusions what price we are paying at the levels of symbolization, relationality, and capacity to tolerate frustration.

Levy (2024) is skeptical: “We know what we are like – grumpy and moody. We have good days and bad days… We lose our patience, we can be violent… We grow old and ugly.” It will not be easy to persuade the broader public to prefer a demanding, ambivalent relationship over a smooth simulation that never gets tired and never refuses.

The remaining hope lies not in turning back the clock, but in preserving spaces where human relationships are still regarded as irreplaceable. Psychoanalytic practice systematically engages with what is painful, ambivalent, and resistant to quick solutions. If it is to remain significant in the era of AI, it will be because it works with what technologies by their very nature bypass. This is not resignation but a call to clarity: we must articulate, with precision and without defensiveness, what we offer that machines cannot. And why that matters.

References

Altman, S. (2025). OpenAI’s Sam Altman on Building the ‘Core AI Subscription’ for Your Life. Sequoia Capital AI Ascent conference, May 2025. YouTube. https://www.youtube.com/watch?v=ctcMA6chfDY

American Psychoanalytic Association Commission on AI. (2024, October 31). ‘Resistance to AI: Does it matter?’ presented by Dr. Amy Levy, Psy.D. [Lecture video]. YouTube. https://www.youtube.com/watch?v=twBatrYNTiI

Bernardi, S. (2025, May 4). AI development – a change in paradigm. International Society of Transference-Focused Psychotherapy. https://istfp.org/ai-development-a-change-in-paradigm

Bion, W. R. (1962). The psycho-analytic study of thinking: A theory of thinking. International Journal of Psycho-Analysis, 43, 306–310.

Essig, T., & Levy, A. (2025). After the symposium: Clinical, ethical, and existential replies to the Q&A generated during the APsA online symposium ‘Artificial Intelligence and our Psychoanalytic Future’. American Psychoanalytic Association. https://apsa.org/ai-psychoanalytic-future-symposium-qa-replies/

Frankfeldt, V. (2025). Beyond the algorithm to emotional communication: Why AI lacks the heart of psychoanalysis. Psychoanalysis, Self and Context, 20(3), 346–368. https://doi.org/10.1080/24720038.2025.2483922

Norton-Poulin, M., & Steiner, V. (2025, May 4). Should we fear artificial intelligence? International Society of Transference-Focused Psychotherapy. https://istfp.org/should-we-fear-artificial-intelligence/

Possati, L. M. (2022). The algorithmic unconscious: How psychoanalysis helps in understanding AI. Routledge.

Singer, N. (2025, August 10). Goodbye, skills. A.I. tech lets students compete for jobs without them. The New York Times. https://www.nytimes.com/2025/08/10/technology/coding-ai-jobs-students.html