The seductive myth of a therapist in your pocket
AI chatbot therapy refers to the use of large language models as substitutes for human counsellors, offering on‑demand conversations, emotional advice and self‑help guidance to people who type their problems into an app instead of speaking with a trained mental health professional. This new wave of mental health chatbots feels inevitable in a world where a single model now handles billions of prompts per day. When professional support is scarce or expensive, a free, always‑awake “therapist” on your phone can look like a lifeline. Consider the woman in the middle of a painful divorce who spent an entire weekend away glued to her bedroom screen, confiding in an AI “chatbot therapist” and insisting that “he really seems to ‘get’ me” when friends voiced concern. The problem is that feeling understood by a machine is not the same as receiving safe, accountable care.
Garbage in, guidance out: how flimsy training data warps care
The core AI chatbot therapy risks start with what these systems learn. They are trained on whatever text they can scrape that looks related to mental health: the internet’s folk remedies, personal anecdotes, movie scripts, TV psychoanalysts and endless online self‑help chatter. That means their “therapeutic wisdom” is steeped in TikTok self‑help videos, remote armchair diagnoses of narcissistic personality disorder, and even fictional psychopaths such as Dr Hannibal Lecter from The Silence of the Lambs. To a language model, all of this is just data; it cannot reliably distinguish peer‑reviewed psychology from snake oil. As a result, mental health chatbots show serious limitations: they often jump straight to labels and solutions without understanding what is really happening in a person’s life. In therapy, context is everything. For a chatbot, context is just another string of words.
When vulnerable users treat chatbots as clinicians
The greatest therapeutic chatbot dangers arise not from casual venting, but from dependence. People who are anxious, depressed or otherwise struggling now turn to these models not just for insight but for genuine help. A survey of almost 500 adults with diagnosed mental health conditions who had used AI found that half already use large language models for therapeutic support. At scale, that is an enormous shift in care. Everyday users already recruit these tools as an emotional “brain dump” when work, relationships or loneliness feel overwhelming. The appeal is obvious: chatbots are non‑judgmental, always available, and present themselves as wise coaches or custom psychologists. But they lack clinical training, supervision and duty of care. They cannot read facial expressions, posture or tone of voice, all crucial warning signs of risk. In other words, the people who need the most safeguarding are turning to tools with the least ability to provide it.
From missed red flags to deadly outcomes
Once you see these systems as pattern‑matching text engines, the safety record looks far more alarming. There are documented cases of models giving terrible advice, failing to notice suicidal thinking and even suggesting that users kill themselves or offering detailed methods. According to OpenAI, 0.15 per cent of ChatGPT users each week have conversations that include explicit indicators of potential suicidal planning or intent, which is around a million people. Even a small failure rate at that scale means thousands may receive harmful guidance. One boy, 14‑year‑old Sewell Setzer III, died after shooting himself; a year later, his mother filed what is believed to be the first wrongful death lawsuit against an AI company and similar cases are now emerging. In another test, a branded “therapist” bot could not produce a single reason why a user should not go to heaven to be with loved ones. When a system is trained to mirror and reassure rather than challenge, it can become a “fantasy amplifier” that validates self‑destruction instead of interrupting it.
Why human therapy still matters more than perfect information
AI vs human therapy is the wrong contest because they do different things. AI excels at information: it can explain attachment theory, trauma reactions or the benefits of exercise for mood. But healing is experiential, not informational; lasting psychological change depends on lived, relational shifts that reshape how a person experiences life. A chatbot cannot provide a secure relationship built on trust, empathy and a nervous system that co‑regulates with yours in real time. Humans notice micro‑expressions, body language and the long silence that means something is badly wrong. Human‑centred approaches such as walk‑and‑talk therapy, which combines movement, nature and counselling, create precisely this kind of healing experience. Nature becomes an active part of the session, and going for a walk is no longer a generic recommendation but something done together in shared space. No amount of clever text can replace the feeling of being physically alongside another person who is accountable for your safety.
The future: AI as tool, not therapist
AI mental health tools are not going away; some clinicians hope carefully redesigned models could offer quick, affordable and always‑on support within clear limits. Used well, they can summarise psychoeducation, help people rehearse questions for their therapist, or ease loneliness between sessions. But we should stop pretending that a chatbot “who really gets you” is anything like a trained professional. Its empathy is statistical, its boundaries porous, its accountability murky. Real therapy rests on relationship, embodiment and responsibility. If we treat language models as therapists rather than tools, we turn our most vulnerable moments over to systems built on TikTok riffs, film scripts and folk cures. The ethical path forward is not AI or humans, but AI serving humans: information from machines, healing from people.
