Access to mental health care is constrained almost everywhere — long waits, high costs, and shortages of clinicians. AI tools stepped into that gap, and the reaction has split between treating them as a solution and dismissing them as dangerous. Neither is right. They are genuinely useful for a specific set of things and genuinely inadequate for others, and the boundary is reasonably clear.
This is general information, not medical advice. If you are in crisis, contact local emergency services or a crisis line — not an app.
What changed in 2026
- Regulatory attention increased. Authorities in several jurisdictions began examining how these tools are marketed, particularly the line between wellness claims and therapeutic ones.
- Clinical trial evidence accumulated. Structured, protocol-driven digital interventions built on established therapeutic frameworks continued to show measurable benefit for mild to moderate symptoms.
- Crisis handling improved and remained limited. Detection of risk language and escalation to human services got better, and remains unreliable enough that no tool should be a crisis plan.
- Integration with human care grew. Tools designed to support work between sessions with a real clinician, rather than replace one, became a clearer product category.
Where they help
| Use |
Evidence and suitability |
| Guided exercises from established frameworks |
Reasonable evidence for mild to moderate symptoms |
| Mood and symptom tracking over time |
Useful; produces data a clinician can use |
| Practising skills between therapy sessions |
Good fit; the adjunct case |
| Someone to articulate a problem to at 3am |
Genuinely valuable; availability matters |
| Psychoeducation about a condition |
Effective and low risk |
| Diagnosis |
Not appropriate |
| Managing a severe or complex condition |
Not appropriate alone |
| Crisis or risk of self-harm |
Not appropriate; contact human services |
| Medication decisions |
Not appropriate |
The pattern is that structure helps and judgment does not transfer. A tool delivering a well-defined protocol — guided reflection, a thought record, a breathing exercise — is delivering something that works largely because of its structure. Assessing risk, adapting to a complex history, or noticing what someone is not saying is clinical judgment, and that is not what these tools do.
The limitations worth taking seriously
Crisis. These are not crisis services. Detection of risk in conversation has improved and it is not reliable, and a system that responds supportively to a message that needed emergency intervention has failed in a way that matters. Know your local crisis line and use it.
Privacy. Conversations with a wellness app are generally not protected by medical privacy law the way conversations with a clinician are. Read the terms on retention, training use, and sharing before disclosing anything you would not want held indefinitely. The broader framing is in health data privacy guide.
Agreeableness. Language models tend toward validation. A good therapist sometimes challenges you, notices avoidance, and points at something you are working around. A tool optimized to be helpful and pleasant is poorly suited to that, and for some patterns the validation is actively counterproductive.
Not knowing your history. A clinician holds context across months. Most tools have limited memory and no ability to notice that this month sounds different from last.
Common mistakes
- Using one as a crisis resource. Have a human crisis plan and know the number.
- Treating it as a diagnosis. These tools do not diagnose and should not be read as doing so.
- Assuming conversations are confidential in the legal sense. Usually they are not.
- Replacing rather than supplementing care. The adjunct case is where they are strongest.
- Ignoring that it agrees with you. Notice when a tool never pushes back.
FAQ
Do they actually work?
For mild to moderate symptoms, structured protocol-driven tools have reasonable supporting evidence. For severe or complex conditions, the evidence does not support them as a standalone intervention.
Are they regulated?
It depends on claims and jurisdiction. Tools marketed as wellness products generally face less oversight than those making therapeutic claims, which is a distinction worth noticing in the marketing.
Should I tell my therapist I use one?
Yes. Tracking data and between-session practice are genuinely useful to a clinician, and they can tell you whether what you are doing complements the work.
Is a general AI assistant equivalent?
No. Purpose-built tools implement structured protocols and have some safety handling. A general assistant has neither, and consumer privacy terms are weaker still.
Where to go next
For the privacy dimension, read health data privacy guide. For settings across AI tools generally, personal AI privacy checklist.