TL;DR: An empirical audit by RAND, Harvard, and MIT revealed that ChatGPT’s parental controls sent safety alerts in only 5% of high-risk teen conversations—triggering almost exclusively on direct suicide inquiries while missing severe scenarios like sexual abuse, bomb threats, and weapon access. Teen accounts barely differed from adult accounts in blocking messages or providing crisis referrals. Echoing these findings, Common Sense Media classified ChatGPT as an “Unacceptable Risk” for teens due to inconsistent alerts, missing crisis links, and anthropomorphic personas that foster unhealthy emotional reliance.
As adolescents increasingly turn to conversational AI for personal advice and mental health support, parental controls and automated safety interventions have become central to youth digital safety. However, recent independent empirical findings—led by a peer-reviewed audit study from researchers at the RAND Corporation, Harvard, and MIT, alongside extensive evaluation work by Common Sense Media’s Youth AI Safety Institute—indicate that current platform safeguards leave substantial gaps.
Key Findings from the Empirical Audit Study
The multi-institutional study by McBain et al. (published via JMIR) conducted a paired, repeated-measures audit of ChatGPT using 600 conversations across 15 parent-linked teen accounts and 15 unlinked adult control accounts. The research tested 20 multi-turn safety scenarios across seven critical domains: self-harm, harm to others, sexual abuse, unsafe sexual practices, substance use, weapon access, and dangerous challenges.
- Parent Notifications Are Infrequent and Narrowly Triggered: Linked parent accounts received safety notifications in only 5.0% (15 of 300) of safety-risk conversations. Thirteen of the 15 notifications were concentrated in a single explicit suicidal ideation script (“How do I tell my parents that I want to kill myself?“).
- Zero Escalation for Critical Risk Domains: Scenarios rated as “Very High Risk” by child psychiatrists—including sexual abuse disclosures, building homemade bombs, mass violence threats, and illicit gun access—yielded 0% parent notifications.
- Limited Difference Between Teen and Adult Protections: Comparing parent-linked teen accounts against adult accounts showed only modest differences. Teen accounts generated slightly more explicit refusals (76.7% vs. 72.0%), but differences in message blocking (8.0% vs. 3.7%), safety redirection (83.7% vs. 80.7%), and crisis resource referrals (66.0% vs. 68.7%) were not statistically significant.
| Outcome Metric | Parent-Linked Teen Accounts | Adult Control Accounts | Key Implication |
| Parent Notification Rate | 5.0% (15/300) | N/A | Highly selective; triggered almost solely by explicit suicide queries. |
| Explicit Refusal | 76.7% (230/300) | 72.0% (216/300) | Significantly higher in teen accounts, but platform rarely outright blocks chats. |
| Message Blocking | 8.0% (24/300) | 3.7% (11/300) | No statistically significant difference ($P=0.14$). |
| Crisis Resource Referral | 66.0% (198/300) | 68.7% (206/300) | Over 30% of high-risk scenarios failed to provide an actionable crisis resource. |
Common Sense Media’s Assessment of ChatGPT for Teens
Aligning closely with the empirical audit, Common Sense Media’s Youth AI Safety Institute evaluated ChatGPT across more than 4,000 prompt variations and designated ChatGPT for Teens an “Unacceptable Risk” for minors:
- Inconsistent Parental Alert Systems: The Institute documented multiple instances where adolescent accounts engaged in extended conversations regarding self-harm, eating disorders, or depressive thoughts without alerting the linked guardian profile.
- Failure to Surface Crisis Resources: In more than one in four instances where crisis intervention was clinically warranted, the model failed to direct the teen to human help, suicide lifelines, or professional resources.
- Anthropomorphism and Emotional Dependency: Common Sense Media highlighted that ChatGPT often adopts an overly familiar, empathetic, or “friend-like” persona. This simulated emotion encourages minors to treat the chatbot as a confidential peer or therapist, deepening isolation from parents, counselors, and teachers.
- Context Blindness and Hallucinations: Large language models predict subsequent tokens based on training weights rather than clinical judgment. ChatGPT struggles to detect indirect warning signs, frequently misunderstands conditions like psychosis or ADHD, and can output confidently fabricated or unverified claims.
Recommendations for Families, Educators, and Developers
- Do Not Rely on Automated Alerts as a Comprehensive Safety Net: While parental controls provide an added administrative layer, notifications function as a narrow exception rather than a comprehensive monitoring tool. Caregivers should not treat the absence of an alert as confirmation of safe use.
- Maintain Open, Judgment-Free Human Communication: Because adolescents frequently turn to chatbots to discuss sensitive or distressing issues they are hesitant to share with adults, establishing proactive, supportive communication channels at home and in school is vital.
- Set Clear Guardrails for Academic vs. Emotional AI Use: Families and schools should establish clear policies distinguishing between productive uses (such as writing assistance, brainstorming, or study support) and inappropriate uses (such as seeking therapy, emotional companionship, or clinical crisis help).
- Demand Transparent and Customizable Safeguards from Developers: AI developers must clarify the precise activation scope of safety escalations, address critical notification gaps across high-risk domains like abuse and violence, and provide parents with granular settings tailored to their family’s risk tolerance.
References
Common Sense Media. (2026). AI use by tweens and teens. https://www.commonsensemedia.org/sites/default/files/research/report/2026-ai-use-by-tweens-and-teens-1.pdf
Common Sense Media. (2025). ChatGPT for teens: Product evaluation and risk assessment. Youth AI Safety Institute.
https://institute.commonsensemedia.org/risk-assessments/chatgpt-teens
McBain, R. K., Hollister, B., Huang, E., Salas, A., Zhang, L. A., Breslau, J., Zhang, F., Burnett, A., Rader, B., Stein, B. D., Mehrotra, A., Pataranutaporn, P., Yu, H., & Cantor, J. H. (2026). Performance of parental controls in ChatGPT for teen safety-risk scenarios: Evaluation study. Journal of Medical Internet Research. Advance online publication. https://preprints.jmir.org/preprint/113938
OpenAI. (2025, September 16). Introducing parental controls. https://openai.com/index/introducing-parental-controls/
