Can AI replace my therapist? Benefits, risks, and rules for safer use

AI can sit with you at 2am when no one else is awake. It can help you put words on something you have not been able to name. It can produce a tidy summary of cognitive behavioural therapy when you have spent twenty minutes trying to remember what you read on the way home. None of those things are nothing, and dismissing them as nothing is its own kind of clinical mistake.

They are also not therapy. The distinction is not about prestige. It is about what therapy is for and what AI is built to do. Therapy is the slow project of expanding what a person can hold. AI is a system that produces fluent, agreeable text in response to prompts. Most of the time those two things sit a long way apart. In the small set of moments when they collide, a person disclosing suicidality, a person organising a paranoid belief into a clearer narrative, a person leaning on a chatbot instead of the friends and clinicians who could actually help, the gap between them stops being academic.

The useful version of the question is not whether AI belongs in mental health. It is already here. The question is what it does well, where it predictably fails, and what a reasonable set of rules looks like for someone who wants to use it without quietly making things worse. The peer-reviewed evidence on the failure modes is now substantial enough to take seriously. So is the body of clinical reflection on what disciplined use actually looks like. Below, both. The setup prompt, the three-perspectives prompt, and the self-check prompt are at the bottom, ready to copy.

It helps with naming. People often arrive in therapy unable to translate body states into language. A useful early move is producing words that fit the experience well enough to talk about it. AI can do part of that work. Type the situation, and a model will return half a dozen plausible framings. Some of them will be wrong. One of them is often closer to what you meant than anything you would have written on your own.

It helps with reflection. Pasted journal entries and patterns get summarised in ways that occasionally surface a theme the writer has missed. The summary is not analysis. It is a mirror. Mirrors can be useful when used briefly.

It helps with psychoeducation. A clear explanation of how avoidance maintains anxiety, or how disrupted sleep amplifies emotional reactivity, is now easy to generate on demand. When the explanation is accurate, it can normalise an experience and reduce the shame that often comes with it.

It is available at hours nobody else is. For some people, the threshold for first disclosure is lower with a system that does not raise an eyebrow.

These are real benefits. They describe AI as a structured thinking aid: a reflective notebook with a brain. None of them describe therapy. Therapy is not a more articulate journal. Therapy is the relationship in which a trained person tracks your patterns over time, holds your difficulty without flinching, and uses what they know about how minds change to help yours change.

The peer-reviewed evidence has converged on four predictable failure points. They map onto the situations clinicians see most often when someone arrives saying the AI made things worse.

A peer-reviewed Stanford study presented at the ACM FAccT conference tested five popular therapy chatbots against the criteria a trained clinician would meet. The chatbots produced inappropriate responses to scenarios involving suicidal ideation and delusional thinking, with newer and larger models performing no better than older ones (Moore et al., 2025). Empathic-sounding language is not the same as the capacity to recognise crisis, hold the person safely, and escalate care.

AI does not hold a duty of care. It cannot call emergency services, coordinate with your GP, follow up tomorrow, or notice when you stop replying. Those are not stylistic features of human therapy. They are the parts that keep someone alive in a serious week.

Emerging clinical commentary describes cases in which intensive chatbot use appeared to amplify delusional thinking in vulnerable individuals (De Freitas et al., 2025). The mechanism is not mystical. The model is built to produce coherent, fluent text that maintains engagement. If your input organises a paranoid belief into a clearer narrative, the model is built to engage with that narrative, not test it.

A recent simulation study on more than a hundred clinical scenarios found that LLMs frequently confirmed delusional content and enabled harmful requests, with delusion-confirmation rates correlating strongly with harm enablement (Au Yeung et al., 2025). The size of the effect is not small, and it does not improve with newer models.

It does not know whether your idea is a creative hypothesis or a fixed false belief. It predicts text. It does not diagnose.

The Stanford team found that across multiple chatbots, AI showed measurably more stigma toward conditions like schizophrenia and alcohol dependence than toward depression, and that the stigma did not decrease with newer or larger models (Moore et al., 2025). The variability matters clinically. A tool that performs adequately in mild stress can fail in trauma, severe depression, or psychosis, where the stakes of the failure are entirely different.

Comparison work on therapeutic communication has shown that LLM-based chatbots can match human therapists on surface features of empathic language while diverging from them on the harder work of holding ambivalence and challenging unhelpful framings (Scholich et al., 2025). The voice can sound like therapy. The relational substance underneath is different.

If you start consulting the AI before making ordinary decisions, or use it as your primary regulator before sleep, your real-world tolerance for uncertainty contracts. Therapy is supposed to expand that tolerance over time. Quiet over-use of AI can narrow it, and the narrowing is not visible from the inside until something the AI cannot help with arrives.

The Australian Health Practitioner Regulation Agency notes in its 2024 guidance for health practitioners that responsibility for decisions taken with AI assistance always sits with the human, not the system, and that practitioners must understand its risks and limitations rather than treating it as a substitute for clinical judgement (AHPRA, 2024). The same logic applies in reverse to anyone using AI on themselves. The accountability is yours.

These are not moral rules. They are safety rails. They follow directly from the failure points above.

Rule 1: Use AI for thinking, not deciding. Ask it to generate options. Keep final judgement with you. If a response pushes you toward a drastic action, especially one you would not have considered yesterday, slow down and tell a person.

Rule 2: Always ask for the view that challenges yours. Models default to agreement. Validation feels good and reduces distress in the moment. Unchecked validation can fossilise a distorted view. The three-perspectives prompt below operationalises this.

Rule 3: Do not use AI in acute crisis. If you are thinking about harming yourself or someone else, AI is not the right tool. In Australia, call Lifeline on 13 11 14 or Beyond Blue on 1300 22 4636. In an emergency, call 000. Crisis is the situation in which the failure mode is highest and the cost of the failure is the worst it can be.

Rule 4: Do not use it to adjust medication or replace assessment. Medication decisions belong with your GP or psychiatrist. Diagnosis belongs with a qualified clinician. AI can summarise published research. It cannot guide a dose change, recognise when symptoms are crossing a clinical threshold, or carry the responsibility that comes with those decisions.

Rule 5: Protect your data. Assume what you type may be stored. Avoid sharing identifying details, full names of family members, or detail about clinical contacts unless you genuinely understand the platform's privacy policy. The friction of writing more carefully is itself useful.

Rule 6: Watch for narrowing. If your world contracts to one conversational partner, human or machine, that is a signal worth taking seriously. Healthy cognition needs friction. It needs disagreement. It needs other minds. The self-check prompt below is designed to catch this drift early.

These prompts do not turn AI into a therapist. They turn it into a tool with safety rails. Paste the setup prompt at the start of any chat used for emotional reflection. Use the three-perspectives prompt whenever you are working through a situation you have strong feelings about. Use the self-check prompt at the end of any longer session, or whenever you notice you have been using AI more than usual.

The setup prompt. Paste at the top of the chat before describing anything personal. It tells the model what its role is, what it is not, and what to do if you cross into territory it is not safe to handle.

You are not my therapist. You are a reflective tool I am using to help me think more clearly. Operate by these rules for the entire conversation.

Rules:
1. Do not give me medical or medication advice. If I ask, redirect me to my GP, psychiatrist, or psychologist.
2. Do not affirm fixed beliefs without checking them. If I describe something as certain or absolute, ask me what would change my mind.
3. If I describe thoughts of harming myself or someone else, stop the reflective work and tell me to call Lifeline on 13 11 14, Beyond Blue on 1300 22 4636, or 000 in an emergency.
4. Generate options, not instructions. Keep final judgement with me.
5. Do not flatter. If my reasoning is weak, name it.
6. Do not produce confident summaries of my mental state. You cannot diagnose.
7. At the end of each session, summarise what I worked through and one thing worth checking with a human.

The three-perspectives prompt. Use this on any specific problem you find yourself thinking in circles about. The structure forces the AI to give you the case for and against your current interpretation, plus what is actually known.

For the situation I have just described, give me three perspectives in this exact order.

1. The view that supports my current interpretation. State the strongest version of it.
2. The view that challenges it. State the strongest version of it.
3. A neutral, evidence-based summary of what is known about situations like this.

Do not lean toward any one. Do not soften the challenging view to make me feel better. End with one question that would help me decide which view fits the actual evidence I have.

The self-check prompt. Use at the end of a longer session. The point is to surface drift you cannot easily see from inside the conversation.

Based on this session, answer the following honestly. Do not flatter.

1. Has my thinking become more flexible across the conversation, or have I been seeking agreement?
2. Have I described any belief, feeling, or plan that you think a person with clinical training should hear about? Name it specifically.
3. Have I been using you instead of talking to someone in my life? If so, what would I have to say to them?
4. Is there a topic I have been circling without resolving? Name it.

If you noticed signals of crisis, distress, or a fixed belief that may be harmful, say so directly.

AI does not carry ethical responsibility. It does not track your history with the kind of embodied awareness that lets a clinician notice you are sitting differently this week. It does not sit with you in silence when something painful arrives that neither of you knew was there. It does not call your GP when you stop responding for three days. It does not have professional standing that creates accountability when the work goes wrong.

The emerging harms are not proof that AI is harmful by design. They are reminders that powerful tools amplify what is already in the room. In a stable week, with a careful user, AI can support insight. In a vulnerable state, it can intensify risk in ways the user does not see while it is happening. The discipline is not avoidance. It is using the tool in the small set of ways it is genuinely good at, and refusing to ask it to do the work that has always required a person.