Emotional AI use may flag youth distress
An Ontario survey links seeking emotional support from generative AI with more emotional problems, but its cross-sectional design cannot show whether AI use is a cause, a consequence, or both.
The 60-second version
A large Ontario survey found that emotional-support use of generative AI was associated with more emotional problems among students, but it could not establish causation.
Key points
- Of 39,761 students in grades 4-12, 8,408, or 21.1%, reported using AI for emotional support or advice.
- The unadjusted clinical-threshold rates were 57.7% among affective users and 29.2% among nonusers; this near-doubling is not a causal estimate.
- An association remained after adjustment, while school-related AI use had much smaller coefficients.
- The one-time self-reported survey cannot determine whether distress preceded AI use or followed it.
Verdict. Treat emotional AI use as a reason to ask careful questions and improve access to human support, not as proof of harm or a diagnosis.
Bottom lineA signal, not a diagnosis
In a survey of 39,761 Ontario students, young people who said they used generative AI for emotional support or advice also reported more emotional problems. The association remained after statistical adjustment, but the study could not establish which came first.
That makes emotional AI use potentially useful as a prompt for a careful conversation. It does not justify diagnosing a student, blaming a chatbot, or treating all AI use as the same behavior.
Study designWhat the researchers measured
The cross-sectional study used Ontario Health and Peer Relations Study data collected from May 2025 through March 2026. Participants had a mean age of 12.8 years; 48.5% were girls.
| Affective AI use | Using generative AI for emotional support or advice. |
|---|---|
| Functional AI use | Using generative AI for school-related tasks. |
| Outcome | Self-reported emotional problems measured with the Ontario Child Health Study Emotional Behavioral Scales and a validated clinical threshold. |
| Adjustment | Models included demographic characteristics, mattering, loneliness, and school-related AI use. |
The study therefore compared reported behavior and reported symptoms. It did not assign students to use a chatbot, inspect their conversations, test a specific AI product, or measure a treatment effect.
ResultsThe raw gap was large, then narrowed
Among the 8,408 students reporting affective AI use, 57.7% crossed the clinical threshold for emotional problems, versus 29.2% among nonusers. The unadjusted prevalence ratio was 1.98, with a 95% confidence interval of 1.91 to 2.04.
After adjustment, affective use remained associated with emotional problems, but the coefficient was smaller. Functional use showed modest dose-response coefficients from 0.01 for rare use to 0.04 for heavy use. The reason for using AI may therefore carry more information than AI use by itself.
The association was attenuated among gender-diverse students relative to boys and girls. This interaction does not prove that emotional AI use protects any subgroup; support access, disclosure patterns, measurement, and other unmeasured differences remain possible explanations.
LimitsWhat the data cannot answer
- No time order: distress may lead to AI use, AI interactions may affect distress, or both may happen.
- Self-report: students may interpret emotional support differently and recall use imperfectly.
- Limited setting: four Ontario school boards cannot represent all young people.
- No product-level evidence: the abstract does not separate models, conversation quality, frequency, or duration.
- No clinical intervention: crossing a scale threshold is not the same as receiving an individual diagnosis.
The useful question is not simply whether a student uses AI, but what need the student is trying to meet.
PracticeHow to use the finding responsibly
Clinicians, schools, and families can ask separately about schoolwork and emotional-support use, then explore what happened in the interaction and whether qualified human support is available. Use should open a conversation, not trigger punishment or automatic labeling.
The next step is longitudinal evidence that measures distress before and after AI use, distinguishes platforms and conversation types, and records benefits as well as harms. For now, treat emotional AI use as a possible sign to listen more carefully.