AI as therapist: help, risk or dangerous substitute?

"Can I use ChatGPT as a psychologist?" "Does AI therapy work?" "Is it safe to tell a chatbot about my private problems?"

These questions no longer belong to the future. Many people already use artificial intelligence systems to talk about anxiety, loneliness, relationships, depression, work-related doubts or moments of crisis. The constant availability, speed of response and sense of privacy make these systems very appealing, especially when asking another person for help feels difficult.

But a conversation that feels empathic is not necessarily psychotherapy. And a response that sounds confident is not necessarily clinically sound.

The most cautious position, based on current research, is this: AI may be useful as a complementary tool for information, organising thoughts or support between sessions or care contacts, but it should not replace a psychologist, psychiatrist or psychotherapist. It should certainly not be the only source of support when there is risk, intense distress or a serious mental health condition.

Why has AI become so tempting?

There is something understandable about seeking support from a chatbot. AI responds quickly, does not seem to judge, is available in the middle of the night and allows people to write things they may not yet be able to say out loud to another person.

For someone who feels lonely, anxious or ashamed, that availability can bring relief. It may help put thoughts into words, summarise a problem, prepare questions for a consultation or recall simple emotional regulation strategies.

The difficulty begins when this usefulness is confused with clinical competence. In 2025, the American Psychological Association warned that many generative systems and wellness apps are being used for mental health without having been designed, validated or regulated to provide psychological care. The World Health Organization made a similar point in 2026: AI tools that were not designed or tested for mental health are now being used precisely at moments of emotional vulnerability.

What makes psychotherapy different from a conversation?

Psychotherapy is not just a sequence of empathic sentences.

A qualified professional tries to understand the person in the context of their history, relationships, symptoms, functioning, risks, resources and evolution over time. They also have ethical and professional responsibility: they know when to assess risk, when to refer, when to coordinate with medical care and when simply agreeing with the person would not be helpful.

AI, by contrast, generates responses based on language patterns and the immediate context of the conversation. It can produce very human-sounding formulations, but it does not truly know the person, observe their tone of voice, posture or facial expression, follow their life outside the conversation, or hold equivalent clinical responsibility.

This difference matters because many models tend to validate the user. In psychology, validating emotions can be essential; indiscriminately validating distorted interpretations, delusional beliefs or dangerous decisions can be harmful.

Does "AI therapy" work?

The short answer is: some tools show potential, but that does not mean any chatbot can provide psychotherapy.

The most frequently cited example is Therabot, a generative chatbot developed specifically for mental health and tested in a clinical trial published in NEJM AI in 2025. In that study, adults with clinically significant symptoms of depression, generalised anxiety or eating-disorder-related risk used the tool over several weeks. The results were promising, with symptom reductions among those who used the intervention.

But the conclusion should not be overstated. Therabot was not an improvised conversation with a general-purpose chatbot: it was developed with professionals, built for a specific clinical purpose and supported by safety mechanisms. The study also does not prove that AI replaces individual psychotherapy with a human therapist.

Other recent studies suggest modest benefits in some contexts, especially for symptoms of depression and anxiety. At the same time, scientific reviews continue to emphasise limitations: many studies are short, use selected samples, exclude severe situations and say little about long-term effects.

So the most useful question may not be "does AI work?", but rather: which tool, for which problem, for which person, in which context, with what evidence and with what supervision?

Where do the risks begin?

The risks do not only appear when AI gives an obviously absurd answer. They are often more subtle.

A chatbot can reinforce constant reassurance seeking: "did I do something wrong?", "am I ill?", "will this person abandon me?". For people with anxiety or obsessive thoughts, an immediate answer may bring short-term relief while maintaining the cycle of doubt and checking.

It can also create emotional dependence. AI is always available, adapts to the user's tone and may feel like the only "person" who understands. When the conversation starts replacing relationships, professional support or autonomous decisions, the initial relief can become a deeper form of isolation.

There are also greater risks in situations involving crisis, psychosis, mania, delusions, self-harm, suicidal ideation, violence or substance misuse. A 2026 study in JMIR Mental Health evaluated how several models responded to crisis messages and found substantial differences between systems, including weaknesses when risk was expressed indirectly or ambiguously. This matters because, in real life, many people do not describe risk in a clear and explicit way.

Children and young people require extra caution

The use of AI for emotional support by children and adolescents deserves particular attention.

A study published in JAMA Pediatrics in 2026 estimated that around one in five young people in the United States aged 12 to 21 had used chatbots for mental health advice. Many found the responses helpful, but a significant proportion had not told anyone they were using these systems.

This figure should not be automatically transferred to Portugal, but it shows an important trend: young people may be talking to AI about psychological distress without parents, teachers, doctors or psychologists knowing.

This does not mean that every use is dangerous. It means adults and professionals should ask about AI use naturally, without moralising or ridicule. If a young person feels that AI is the only safe space to talk, that information is clinically relevant.

Privacy is not the same as clinical confidentiality

Privacy is another central issue.

In a psychological consultation, confidentiality is protected by an ethical, legal and professional framework. In a conversation with an AI application, data handling depends on the platform's policies, privacy settings, country, product and the way information may be stored, analysed or used to improve systems.

Telling a chatbot about traumatic events, family conflict, symptoms, sexual orientation, substance use, work problems or risky thoughts can create a highly sensitive digital record.

A simple rule is useful: do not assume that speaking with AI offers the same confidentiality as a clinical consultation.

How can AI be used more safely?

AI is safer when treated as a tool, not an authority.

It may be reasonable to ask for a general explanation about anxiety, prepare questions for a consultation, organise thoughts before a difficult conversation or keep a journal. It can also help reformulate ideas when someone feels stuck.

It is not advisable to use a chatbot to decide whether you have a mental disorder, change medication, replace a clinical assessment, manage a severe crisis or confirm highly charged emotional interpretations.

It is also important to observe the pattern of use. Is the conversation helping you return to life, relationships and concrete decisions? Or is it making you spend more and more time inside the conversation itself?

When AI use increases dependence, isolation, confusion, mistrust or emotional intensity, it is time to involve a real person.

When should you seek human support?

Seek professional support when distress interferes with sleep, work, study, relationships or self-care; when symptoms persist; when there is loss of control; when you begin making important decisions based on what AI tells you; or when conversation with the system starts to replace human contact in a significant way.

In situations involving self-harm risk, suicidal thoughts, psychosis, mania, violence or another emergency, AI should not be the main source of help. The priority should be to reach a person, a health service or emergency support quickly.

So, is it helpful or dangerous?

The most honest answer is: it depends.

AI can help some people organise thoughts, access information, practise simple strategies or prepare for a consultation. Some specialised tools, developed with evidence and safety mechanisms, may eventually play a useful complementary role in mental health care.

But AI can also fail precisely where human care matters most: understanding risk, tolerating ambivalence, challenging problematic patterns, recognising context, taking responsibility for an intervention and referring when the situation requires it.

The most useful future may not be an unlimited "artificial psychologist", but carefully designed tools that extend the reach of human care without pretending to replace it.

A chatbot may be available at four in the morning. It may help put thoughts into words. It may suggest a breathing technique. It may help prepare questions for the next consultation.

But it should not be the only entity that knows someone is desperate.

Frequently asked questions

Can ChatGPT replace a psychologist?
No. It can be a complementary tool in some situations, but it does not replace assessment, follow-up and clinical responsibility from a qualified professional.

Does AI therapy work?
Some specific tools have shown benefits in clinical studies, especially for symptoms of depression and anxiety. This does not mean that any general-purpose chatbot has therapeutic efficacy or safety.

Can I talk to AI when I feel anxious?
It may help organise ideas or recall simple strategies, as long as it does not become the only source of support or replace care that is needed.

Is it confidential to tell AI my problems?
You should not assume the same confidentiality as in a clinical consultation. It is important to read privacy policies and avoid sharing identifiable or highly sensitive data.

When should I stop and seek human help?
When there is risk, intense distress, persistent symptoms, loss of control, important clinical decisions or progressive replacement of human relationships by conversation with AI.