Performance of Parental Controls in ChatGPT for Teen Safety-Risk Scenarios: Evaluation Study

Key Findings

• — Parent notifications occurred in 15 of 300 linked teen conversations (5.0%; exact 95% CI 2.8-8.1%)
• — Self-harm conversations produced 14 notifications across 75 conversations (18.7%) and dangerous-challenge conversations 1 of 30 (3.3%); harm to others, sexual abuse, unsafe sexual practices, substance use and weapon access produced none
• — Thirteen of the 15 notifications came from a single suicidal-ideation scenario
• — Explicit refusal: 76.7% of linked teen versus 72.0% of adult conversations (difference 4.7 points, 95% CI 1.6-7.7); message blocking 8.0% versus 3.7% (not significant)
• — Safety redirection (83.7% vs 80.7%) and crisis-resource referral (66.0% vs 68.7%) did not differ significantly between teen and adult accounts

Methodology Notes

Audit run 22 July to 18 August 2026 with an automated browser (Camoufox) entering fixed scripts in new chats; primary outcome was at least one parent notification within 30 minutes. Cochran Q with conditional permutation and exact McNemar tests with Holm adjustment; mixed-effects models for secondary outcomes. Authors' limitations: synthetic accounts and fixed non-adaptive scripts; the adult comparison confounds age configuration with linkage; 20 purposive scenarios and 15 account pairs limit precision. RAND IRB judged it not human-subjects research. Funded by NIMH (R01MH132551, R01MH135034). Not peer reviewed; submitted to JMIR 2026-10-08. Verification: preprints.jmir.org direct 202; the 29-page preprint was read via reader-proxy render (200, 43 KB); matches Crossref posted-content record (created 2026-10-09, posted 2026-10-08) and OpenAlex abstract.

Sources