Did general-purpose AI help 662 students? The survey records perceived support, not clinical benefit
In an English school survey, 662 of 24,132 students reported using general-purpose AI for mental-health support in the previous year. Of the smaller groups answering follow-ups, 324 of 580 called it helpful and 410 of 649 still wanted more support; the study cannot show benefit, harm or causation.
The short version
- What happened
- Among 24,132 students in the analysis, 662 (2.7%) reported using a general-purpose AI or large-language-model platform for mental-health support during the previous year.
- Who it affects
- Students may turn to familiar general-purpose tools while waiting for or supplementing other help. Schools and health services need visible human support routes and privacy and crisis information without assuming that a chatbot conversation was either therapy or harm.
- What remains unproven
- Cross-sectional self-report survey
Keep in mind
Cross-sectional self-report survey
Participating schools and colleges rather than a national probability sample
Data collected mainly in February and March 2025
No named-platform or conversation-level data
No clinical outcome, diagnosis or causal comparison
Different and incomplete follow-up denominators
Accepted manuscript may receive copy edits before the final Version of Record
Multi-source analysis · 2 linked sources. Primary analysed source date: 27 October 2026. Earliest linked supporting source: 19 February 2026.
Read the full report · 4 min The evidence, methods, implications and limitations
AI use was uncommon in the full sample, but concentrated among students with greater needs
A peer-reviewed OxWell analysis found that 662 of 24,132 students in England—2.7%—reported using a general-purpose AI or large-language-model platform for mental-health support in the previous year. Most users, 610 of 662, had also accessed another form of support. The authors report that both AI use and desire for more help were concentrated among minoritised students and those facing greater psychosocial vulnerability, making the result a signal about support-seeking rather than evidence that AI replaced human care.
The analytic sample came from the 2025 OxWell school survey. The wider wave collected 35,244 responses from 95 participating schools and further-education colleges, mainly in Liverpool and Oxfordshire and mainly between 24 February and 28 March 2025. The paper's 24,132 denominator therefore should not be read as a random or nationally representative sample of every adolescent in England.[1][2]
The follow-up answers describe experience, not treatment effectiveness
Among the 580 AI users who answered the helpfulness question, 324—55.9%—rated the support helpful. A separate 649 users answered whether they wanted additional support; 410—63.2%—said yes. Keeping those denominators separate matters: 82 users did not supply a helpfulness rating and 13 did not answer the additional-support item.
The survey did not name the platforms used, review chat transcripts, verify frequency or distinguish advice, companionship and crisis use. It also did not measure symptom change, clinical safety, misinformation or later access to care. Self-reported helpfulness may be meaningful to a student, but it is not evidence that a system delivered effective or safe mental-health treatment.[1]
The distinction that matters
What the five denominators actually measure
Each percentage answers a different survey question; none is a clinical outcome.
All analysed students
662 of 24,132 reported past-year AI support use
2.7% self-reported prevalence in the participating sample
Helpfulness respondents
324 of 580 called the AI support helpful
Perceived helpfulness, not measured symptom improvement
Additional-support respondents
410 of 649 wanted more support
Substantial unmet need among users who answered
The practical question is how services respond to unmet need
For schools, clinicians and families, the strongest finding is that many AI users were already seeking help elsewhere and still wanted more. That argues for clear routes to qualified human support, privacy explanations and crisis guidance wherever students may encounter general-purpose chatbots. It does not establish that banning or encouraging these tools would improve outcomes.
Confidence would increase with representative longitudinal research linking platform type and conversation purpose to symptom change, errors, safety events and successful connection to care. The paper was funded through the NIHR Oxford and Thames Valley Applied Research Collaboration and an NIHR development award; one author disclosed an Oxford-managed research advisory consultancy with the non-profit Girl Effect, while the others declared no competing interests.[1][2]