AI chatbots for mental health: benefits and risks

Millions of people now use some form of AI chatbot when they feel anxious, low or simply unable to sleep. Some use a purpose built wellbeing app. Most just open whichever general assistant is already on their phone. The technology arrived faster than the guidance, so this article gives you the honest version: what these tools help with, what the research actually found, and the risks that rarely make it into the marketing.

1. The three kinds of AI people use

Lumping everything together is the first mistake. There are three very different categories.

General assistants. Broad language models that answer anything. Enormously capable, no clinical design, no duty of care, no crisis protocol beyond generic safety messaging.

Purpose built mental health chatbots. Structured tools, often designed with clinicians, that deliver cognitive behavioural or acceptance based exercises in conversation. Narrower, safer, usually less impressive to talk to.

AI companions. Products designed for relationship and personality rather than clinical outcomes. These are the ones people become attached to, and they carry the least clinical oversight of the three.

2. What the evidence shows

The honest summary is modest but real. Reviews of trials involving purpose built conversational agents find small to moderate short term reductions in symptoms of anxiety and depression compared with doing nothing. The strongest results come from tools that deliver a recognised therapeutic method rather than open conversation.

Three caveats matter. Most studies are short, often four to eight weeks. Benefits tend to fade once people stop using the tool. And the comparison group is usually a waiting list rather than actual therapy, so the finding is closer to better than nothing than to as good as a therapist.

There is almost no long term evidence, and very little on people with severe or complex needs, because those groups are usually excluded from the trials.

3. The real benefits

Availability. Distress does not schedule itself between nine and five. Having something to do at two in the morning has genuine value.

Lower shame. Many people write things to a chatbot they have never said aloud. That first articulation can be the start of the change.

Skills practice. Thought records, reframing, breathing, grounding, behavioural activation. These are teachable techniques and a conversational format suits them.

Structure when overwhelmed. Turning a spiral into a short list of next steps is a legitimate help. If your evenings are the hardest part, our guide on how to stop overthinking at night covers the same ground in more depth.

A rehearsal space. Practising a difficult conversation before you have it. Useful alongside our guide on setting boundaries at work.

4. The risks worth knowing

Reassurance loops. This is the risk clinicians worry about most. Compulsive reassurance seeking maintains anxiety, and a tool that answers every time, instantly, makes the loop effortless to feed. Ten reassuring answers feel better in the moment and leave the anxiety stronger.

Confident inaccuracy. A fluent answer is not a correct one. Advice about medication, diagnosis or risk should never be taken from a model without professional confirmation.

Unreliable crisis detection. Safety filters look for explicit risk language. People in real distress often use flat, vague or hopeless language instead. Do not rely on software to notice that you are in danger.

Avoidance dressed as coping. If talking to AI becomes the reason you never message a friend or book a session, the tool is deepening the isolation it appears to relieve.

No accountability. A licensed professional carries training, supervision and ethical obligations. A product's terms of service usually disclaim responsibility entirely.

5. Privacy and your data

Assume less privacy than the intimacy of the conversation suggests. Depending on the provider, your messages may be retained, reviewed by staff for quality, used to train future models, or exposed in a breach. General assistants are usually not health services and are typically outside the scope of health privacy law.

Practical steps: read the retention policy, disable training on your data if that setting exists, avoid naming your employer, partner, children or clinician, and never paste medical records into a general chat window.

This is why anonymity should be designed in rather than promised afterwards. On BYOU the Ask for help feature is anonymous by default. Your name and contact details are never shared with the experts who respond, and the answers come from real qualified professionals across coaching, therapy, nutrition, breathwork and bodywork.

6. AI companions and loneliness

Companion apps deserve their own paragraph because they are the fastest growing category and the least studied. They are warm, always available and incapable of rejecting you. For someone who is isolated, that is a powerful combination.

Short term, they can genuinely soften a hard evening. The problem is what they cannot provide. Research on loneliness suggests that what reduces it is reciprocity, shared activity and being needed by someone. A companion that exists only to attend to you offers none of those. If you are using one daily, treat it as a signal rather than a solution, and read our article on asking for help.

7. A safety checklist

• Give it one job. Decide what you want before you open it, and close it when you have that.
• Keep sessions short. Long late night sessions tend to be rumination, not support.
• Verify anything clinical. Diagnosis, medication and risk belong to professionals.
• Notice the loop. Same question, new words, means stop and call someone.
• Protect your data. Fewer identifying details, always.
• Keep one human in the loop. A friend, a colleague, a professional, weekly.
• Know your escalation route. Save a crisis number now, not later.

8. What to use instead when it matters

If your difficulty has lasted more than two weeks, affects your sleep, work or relationships, or involves trauma, self harm, disordered eating, substance use or abuse, a chatbot is the wrong tool. You need a qualified person, and the barrier to that is lower than most people assume.

On BYOU you can book a single session with a verified therapist, coach or specialist and pay only for that session. Every expert sets their own price and you see it before you book. There is no subscription and no minimum. If you are unsure who to speak to, our comparison of therapy and coaching will help you choose.

If you are in immediate danger, contact your local emergency number or a crisis line. In the United States call or text 988. In the United Kingdom call 116 123. In the Netherlands call 113 or 0800 0113.

Used with clear limits, AI can be a useful part of looking after your mind. It works best as the thing that gets you to the door, not the thing waiting on the other side of it.