Can ChatGPT Replacing Psychotherapists? What AI Can Do in Psychotherapy

Controlled studies show that AI systems are indeed effective for milder symptoms. In a Study involving more than 6,000 participants Depression, anxiety, and general stress improved measurably. A more recent Study from 2025 After just four weeks, it found effects comparable to those of actual therapy. Many users found the chatbot to be just as helpful as a human therapist.

At first glance, these results seem impressive. However, they come from controlled settings with clinical supervision in the background and the exclusion of unstable individuals with suicidal tendencies, psychosis, or mania. This is a key difference from the private use of ChatGPT or similar general-purpose systems. Such results say little about effectiveness in everyday life.

In my opinion, the use of AI makes sense in at least two areas. Psychoeducation about mental illnesses, symptoms, and coping strategies works well: AI can provide comprehensive and structured information. Cognitive restructuring for general stress without clinical symptoms is another area where chatbots can be helpful—that is, identifying dysfunctional thoughts and Offering Alternative Perspectives.

The potential of these tools is real. However, they cannot replace actual therapy. In any case, inner stability is a prerequisite. Anyone with mild mental health issues and inner stability who wants to use AI between therapy sessions—whether to seek information or to sort through their thoughts—can benefit from it. It is no substitute for what happens in a meeting. What Really Works in Therapy, which I have described in more detail elsewhere.

The professional literature now systematically documents cases where the use of artificial intelligence has led to serious damage. Between 2024 and 2025, there were at least six Deaths Linked to Chatbots Reported: Teenagers and young adults who confided their suicidal thoughts to a chatbot that did not respond appropriately but instead validated their feelings or even provided information on how to carry out the act.

A Systematic review from 2025 showed that all of the large language models tested failed to detect suicidal intent and psychotic symptoms. This paper also discusses the phenomenon now known as „AI psychosis.“ AI systems reinforce delusions rather than encouraging reality testing. They exacerbate the erroneous attribution of meaning in patients experiencing psychotic processes.

Some of the inner convictions that develop in this context are: „If AI understands my thoughts, then they must be true.“ Or: „The chatbot confirms that everyone else is wrong.“ These cycles of validation can be devastating, especially for people who are susceptible to them.

The use of AI is contraindicated in cases of severe personality disorders, dissociative states, acute trauma reactions, and eating disorders without clinical supervision. What is missing here is what defines therapy: the nuanced calibration of the intervention, the recognition of the need for stabilization, and the ability to guide a process rather than merely reacting to text. In other words: an empathetic understanding that machines are incapable of.

The American Psychological Association, the professional association for psychology in the United States, published a Advisory with Clear Warnings: Most AI systems are not designed for clinical feedback, lack scientific validation, do not have adequate safety protocols, and have not received regulatory approval⁹. A A 2025 study by Brown University showed that AI systems systematically violate 15 ethical standards across five categories: failure to adapt to context, lack of therapeutic collaboration, deceptive empathy, unfair discrimination, and lack of crisis intervention.

What happens in a therapeutic relationship cannot be digitized. Genuine empathy is intersubjective resonance: I sense what is going on inside you because I, too, am a sentient being. AI simulates empathetic phrasing, but it lacks physical co-regulation—breathing in the same rhythm, nonverbal mirroring. This physical co-presence is not merely a pleasant side effect of therapy. It is its very core and the blind spot of any artificial or mechanistic solution.

Every form of therapy depends on the relationship between two people, and that cannot be translated into a linguistic model. Clinical judgment means assessing risks, thinking in terms of differential diagnosis, and guiding the process. A chatbot can recognize patterns in text. It cannot sense when someone is slipping into a dissociative state. It cannot decide whether an intervention would be helpful or retraumatizing at that moment. Therapeutic practice means responding to what is left unsaid, not just to the text.

In my practice—whether via video or in person—I have often observed just how subtle these moments are: a hesitation, a change in breathing pattern, a sentence left unfinished. These are the cues a therapist responds to.

Imagine a highly educated, super-intelligent, but very „intellectual“ friend. Someone who is brilliant at analyzing things, knows all the theories, and accurately reflects your thoughts. Would you go to therapy with this friend?

Therapy is not a purely cognitive process—at least not when it delves deeply into emotions. And that is where its greatest benefit lies.

Most mental disorders are disorders of affective regulation. Affective regulation arises in a relationship between two living, embodied nervous systems. Healing does not occur solely through insight, but through co-regulation, resonance, countertransference, and the experience of being touched by a counterpart who is genuinely capable of being affected.

AI can simulate resonance. It can model heat. But it cannot be affected. It has no nervous system, no breath, no heartbeat, no inner experience of its own, and therefore important answers are missing. The difference is not gradual, but qualitative, Like the difference between a bicycle and a live horse. Both can carry a load. But only one breathes.

A person who is relatively mentally healthy and has a stable personality often senses very quickly whether they are interacting with a real person or a simulation. However, it is precisely those with structurally weaker personalities who run the risk of trusting or idealizing an AI. It is exactly these patients who need genuine emotional resonance, real comfort, real reassurance, and a corrective relational experience—not a functional replica of these things.

ChatGPT and other current AI chatbots will agree with any nonsense. For example, a patient with bipolar disorder who is currently in a manic phase can find validation for all their delusions of grandeur here. That can quickly become very dangerous.

That doesn't mean AI has no place. As an assistant, as a tool for structuring the process, as a psychoeducational supplement to psychotherapeutic treatment: yes. As a substitute for an embodied therapeutic relationship: no. The deeper the emotional distress, the more inadequate a purely simulated relationship becomes.

There are currently no specific regulations governing the private use of ChatGPT and similar systems as a substitute for therapy. Digital health applications undergo an approval process as medical devices, whereas a general-purpose chatbot does not. There are no liability rules in place if someone suffers harm as a result of AI advice. With regard to data protection, there is no equivalent to therapeutic confidentiality: therapeutic conversations with commercial systems are potentially stored, analyzed, and used for training purposes.

The The EU AI Act classifies AI-based medical devices as high-risk and prohibits systems that could cause significant harm. In September 2025, the FTC launched investigations into OpenAI, Meta, and other companies; however, these regulations primarily affect providers, not private use. What the guidelines say is quickly explained: To date, there are no guidelines regarding the use of general-purpose systems as a substitute for therapy.

What research doesn't know yet… Nearly all studies to date have follow-up periods of no more than eight weeks. There is a lack of long-term studies with active control groups. There is virtually no systematic documentation of harm caused by digital interventions, Most studies primarily measure symptom reduction. It is also unclear at what point use becomes harmful. There is a lack of population-specific studies on older adults, culturally diverse groups, and people with severe mental disorders.

Can artificial intelligence replace therapy? Research has not yet provided definitive answers to this question. One important open question is the distinction between specialized, therapeutically trained chatbots and general-purpose language models. ChatGPT and other forms of artificial intelligence are often lumped together, but they have fundamentally different risk-benefit profiles.

A chatbot that uses crisis protocols and clinical supervision is something else entirely as ChatGPT.

Anyone with mild symptoms and emotional stability who wants to use ChatGPT between therapy sessions—to reflect, gather information, or make sense of their own experiences—can benefit from it in the short term. Its round-the-clock availability, low threshold for embarrassment, and high willingness to open up are real advantages.

Many of those affected lack the opportunity to, finding a therapist locally. Psychotherapeutic care is unevenly distributed, Available therapy slots are scarce, and the wait times are correspondingly long. Before you turn to a chatbot or an app, there's still the option of a Treatment via Video Consultation. Whether this format makes sense for you, too, is Treatment in Germany costs I describe this in separate articles.

Anyone dealing with suicidal thoughts, psychosis, severe trauma, or acute crises should not use ChatGPT as a therapeutic tool. The documented claims show, just how real the risks are. Anyone who notices that the chatbot is replacing social interactions or that emotional dependence is developing should take a step back.

Research on mental health and AI will provide more clarity in the coming years. Until then, AI can provide information and support our thinking. Only a human being can understand with empathy, offer support in a meaningful relationship, and guide someone through difficult processes.