How AI is changing mental health care

Every few years a technology arrives that is supposed to fix mental health care. Teletherapy, apps, wearables, and now artificial intelligence. Each one changed something real and none of them solved the underlying problem, which is that far more people need support than there are trained people to provide it. This article looks at what AI is genuinely doing to mental health care right now, what is still marketing, and what it means for you.

1. The problem AI is being asked to solve

The gap is enormous. Across most countries, a large majority of people who meet criteria for a common mental health condition receive no treatment at all. Where services exist, waiting times are often measured in months. Clinicians spend a substantial share of their working week on documentation rather than on people. Cost, stigma and geography exclude many more.

Any honest assessment of AI in this field has to be measured against that gap, not against a perfect service that does not exist anywhere.

2. Where AI is already useful

The most valuable uses today are unglamorous and mostly invisible to clients.

Documentation. Drafting session notes and summaries. This is the single largest time saving, and time saved on paperwork is time returned to clients.

Intake and triage. Structured questionnaires, first contact forms and routing someone toward the right kind of professional rather than the first available one.

Between session support. Reminders, exercises and psychoeducation while someone waits for or continues treatment.

Training and supervision. Simulated practice conversations for trainees, and pattern review to help supervisors spot where a clinician might need support.

Language and access. Translating materials and making information available to people who were previously excluded by language.

Matching. Helping someone find a professional who fits their need, budget and availability. This is the part BYOU focuses on, and our how it works page explains the process in full.

3. Risk prediction and its limits

One of the most discussed applications is predicting who is at risk of crisis using patterns in records, language or phone use. The statistical performance of these models is real but modest, and the practical problems are large.

Because crises are relatively rare, even an accurate model produces many false alarms. Acting on all of them is impossible; ignoring them defeats the point. Models also tend to perform worse for groups underrepresented in the training data, which are frequently the groups already receiving the least care.

Used well, a risk flag is a prompt for a human to make contact. Used badly, it becomes an automated judgement about a person who never consented to being scored.

4. The bias problem

AI learns from historical data, and mental health data carries the history of who was believed, who was diagnosed and who was medicated. If certain groups were historically underdiagnosed for depression and overdiagnosed for other conditions, a model trained on those records inherits the pattern and presents it as objective.

Language adds another layer. Distress is expressed differently across cultures, and in many languages it is described physically rather than emotionally. A system tuned on one way of talking about suffering may simply fail to recognise another.

This is fixable, but only with deliberate work: diverse training data, evaluation broken down by group, clinician oversight and a willingness to publish the failures alongside the successes.

5. What this means for waiting lists

AI can genuinely shorten the administrative parts of the pathway. Faster intake, less paperwork, better routing and useful support while waiting all add capacity at the margins.

What it cannot do is manufacture clinicians. The shortage is a workforce issue, and the risk is that software becomes a way of appearing to address demand without funding the people who meet it. Being offered an app when you asked for a therapist is not access.

Marketplaces help differently. By making it easy to find and book an individual professional directly, they open capacity that already exists but is hard to discover. On BYOU you book a single session with a verified expert, every expert sets their own price, and you see it before you book. No subscription and no waiting list.

6. What it means for therapists and coaches

The job is changing shape rather than disappearing. Less time on notes, more time with people. Better preparation before a first session because intake information arrives structured. More competition on discoverability, since clients now research extensively before choosing.

There are new obligations too. Clients should be told when AI is involved in their care. Identifiable client information should not be pasted into general tools. Clinical judgement cannot be delegated to software, no matter how convincing the output. If you work in this field, our page for experts explains how BYOU handles discovery and pricing.

7. What it means for you as a client

Three practical implications.

Ask who is answering. If a service offers support, find out whether a human is involved and at which points. Vagueness here is itself an answer.

Read the data terms. Mental health information is among the most sensitive data you will ever share. Know whether conversations are stored and whether they train models.

Use AI for preparation, humans for the work. Organising your thoughts, learning a technique or drafting the first message are good uses. Being understood is not something software can do for you. Our article on whether AI can replace therapy goes deeper on this distinction.

8. How to judge a mental health app

• Named clinical involvement. Real people with real credentials, not an anonymous advisory board.
• Published evidence. Studies, not testimonials.
• Plain language privacy policy. If you cannot tell what happens to your data, assume the worst.
• Clear crisis guidance. Visible before you need it, not buried in settings.
• Honesty about limits. Trustworthy tools say what they do not do.
• A route to a human. The most important feature of any digital mental health product.

9. The next five years

The most likely direction is blended care. AI handles intake, admin, reminders and between session material. Humans do the relational work: the noticing, the challenging, the being present. That combination could meaningfully improve access without hollowing out quality.

The risk is the cheaper version, where cost pressure quietly replaces people with software and calls it innovation. Which version arrives is a matter of choices being made now by services, regulators and platforms, not an inevitable property of the technology.

Our position is straightforward. Technology should make it faster and less intimidating to reach a qualified person, then step aside. That is why the Ask for help feature in the BYOU app is answered by real professionals rather than a model, and why every reply comes without your name or contact details attached.

If you have been putting off speaking to someone, the useful step is rarely another app. It is one conversation with one person who knows what they are doing.