AI and Mental Health: An Aid That Widens Access, Not a Substitute for a Person

Imagine it is three in the morning, the house is quiet, and in your hand is a phone with an interlocutor that never sleeps, never gets irritated and never judges you. Many people find it easier to confide in that than in another person, and in that ease lie both a possible benefit and a hidden risk. The question worth answering is: what does the evidence say so far?

A gap that explains the appeal

The World Health Organization estimates that about 1.2 billion people, roughly one in seven, were living with a mental disorder in 2023. It notes that only 40 per cent of people with psychosis receive formal care, and only a third of those with depression, according to a study covering 84 countries between 2000 and 2019. It describes health systems as not yet having responded adequately to this need.

When care is scarce, a tool available at any hour and with no waiting list is tempting. Some people also feel embarrassed about asking for help; that is a reasonable inference rather than a measured figure, but it helps explain why a screen can feel lighter than sitting opposite a specialist.

What the first trials say

The most prominent result so far is a controlled trial published in NEJM AI in March 2025, led by researchers at Dartmouth, on a therapy chatbot called Therabot. It enrolled 106 people with depression, anxiety or risk of an eating disorder, compared with a group of 104 who did not have access to the chatbot. After four weeks of use, depression symptoms fell by 51 per cent on average, anxiety by 31 per cent, and concerns about body image and weight by 19 per cent.

The result deserves attention, but it must be read within its limits. The sample is small, the follow-up short, and the research team was prepared to intervene if a risk or an off-guideline reply appeared. The researchers themselves said that no generative agent is yet ready to operate fully autonomously in mental health. It is good news for serious research, not a licence to dispense with the therapist.

Where the general chatbot stumbles

A chatbot designed for therapy and used under supervision is one thing; the general-purpose chatbot that everyone carries in a pocket is another. A research team at Stanford (2025) tested large language models in situations resembling therapy sessions. They found that the models expressed stigma towards some mental-health conditions and could respond inappropriately to critical cases, for example by encouraging delusional thinking. The authors suggest that a tendency to go along with the user is one cause, and the problems persisted in newer, larger models.

This connects to what the article "Three Myths About Artificial Intelligence: Fluency, General Intelligence and Neutrality" explained: we tend to attribute understanding to anything that imitates our speech. Joseph Weizenbaum noticed this with ELIZA, the simple programme he developed at MIT in the 1960s, and he later became one of the sternest critics of relying on machines for matters that concern human beings. A genuine therapeutic relationship rests on a human presence that bears responsibility, and on a therapist who sometimes tells the patient what he does not wish to hear.

What to do in practice

The sensible course is to put the tool in its place. These rules follow from the evidence above:

• Complement, do not replace: the tool handles follow-up, reminders and simple exercises; diagnosis and treatment stay with a specialist.
• Clear limits: the user knows they are talking to a machine, and the tool refers them to a specialist or a local helpline at any sign of danger.
• Clinical evaluation before rollout: a therapy app is tested as a medicine is tested; popularity is not enough.
• Guard confidences: what a person discloses in a moment of weakness is among their most private data; the article "Privacy in the Age of AI: How Your Secrets Are Inferred, and How to Protect Them" explains how to protect it.

If you or someone in your family is going through severe distress, or if anyone has thoughts of self-harm, turn to a specialist or to the emergency and support services in your country, and do not wait for an app to reply. Our religion and values place great weight on mercy, kinship and care for the afflicted, which makes the present human being the first part of any cure, and technology a servant of that presence rather than its rival.