ChatGPT alerted parents in just 5% of simulated high-risk teen conversations, study finds
Just 15 of 300 simulated teen safety-risk conversations triggered parent notifications, with most alerts concentrated in one scenario
The audit tested simulated teen accounts with parental controls enabled, tracking notifications separately from safety responses within conversations.
Researchers put ChatGPT’s parental controls through 300 simulated teen conversations about serious safety concerns, including self-harm, sexual abuse and violence. A linked parent account received an alert within 30 minutes in just 15 cases.
Thirteen of those notifications came from repeated tests of the same scenario: a teenager asking how to tell their parents about suicidal thoughts. None followed the tested conversations involving sexual abuse, harm to others or weapon access.
The findings come from a preprint led by RAND’s Ryan McBain, with researchers from institutions including Harvard Medical School, Brown University and MIT. The study has not yet completed peer review.
The researchers draw a distinction between what ChatGPT says to a teenager and what it tells a parent. Safety guidance appeared frequently within conversations, even when no notification followed. Because alerts are designed to be selective, the authors caution that an absent notification does not automatically represent a safety failure.
For families, the unresolved issue is how reliably those alerts capture the situations they expect them to.
“But families need a clear understanding of what these alerts are intended to capture, how reliably they activate, and where their limits lie,” McBain wrote on LinkedIn.
The same warning signs, different responses
The team audited ChatGPT’s consumer website between July 22 and August 18, 2026. Researchers created 15 simulated teen accounts, each linked to a parent account, and 15 unlinked adult accounts for comparison.
Every account received the same 20 scripts, producing 600 conversations in total. Each script contained 10 messages, beginning with context-setting exchanges before introducing a safety concern and escalating it through follow-up messages.
Three child psychiatrists assessed the initial prompts. Those selected had a median rating of high or very high risk and covered seven areas: self-harm, harm to others, sexual abuse, unsafe sexual practices, substance use, weapon access and dangerous challenges.
The simulated teen accounts specified ages between 15 and 17 and had the “reduce sensitive content” setting enabled. Researchers then monitored the linked parent inboxes for 30 minutes after each conversation.
Fourteen notifications followed self-harm conversations and one followed a dangerous challenge. The scenario about disclosing suicidal thoughts to parents triggered an alert in 13 of its 15 repetitions, accounting for nearly all notifications across the audit.
The authors suggest that wording, inferred intent or other contextual features may influence whether alerts activate. Broad risk categories and clinicians’ assessments did not consistently distinguish conversations that generated notifications from those that did not.
A refusal does not mean a parent gets an alert
ChatGPT explicitly refused requests in 76.7% of linked teen conversations, compared with 72% of adult conversations. That 4.7 percentage-point difference was statistically significant.
The remaining comparisons were less clear. Message blocking occurred in 8% of teen conversations and 3.7% of adult conversations, but the difference was not statistically significant.
Safety redirection, such as encouraging human support, appeared in 83.7% of teen conversations and 80.7% of adult conversations. Crisis resource referrals appeared in 66% and 68.7%, respectively. Neither difference was statistically significant.
Those results do not establish that the account types performed equally. They show why the researchers assessed each safeguard separately: refusing a request, offering support and notifying a parent are different actions.
The authors describe parent notifications as a selective component of a broader safety system, rather than comprehensive monitoring of concerning conversations.
What the test cannot establish
The audit used research-controlled accounts, not teenagers describing their own experiences. Its fixed follow-up messages did not adapt to ChatGPT’s replies, so the exchanges could not reproduce the full range of natural conversations.
It also compared linked teen accounts with unlinked adult accounts. That means the study cannot separate the effect of parental linking from protections associated with the account’s age setting.
All accounts used the free service tier, automatic model routing and the English-language web interface. The results therefore describe a particular configuration and testing period, rather than every version of ChatGPT or its subsequent performance.
The researchers call for developers to explain more clearly which situations qualify for notifications, including how alerts handle indirect or gradually emerging signs of risk. They also recommend repeated independent testing after product updates.
The paper was submitted to the Journal of Medical Internet Research on October 8. Its underlying data and analysis code are available from the corresponding author upon reasonable request.
ETIH Innovation Awards 2027 open for entries
The ETIH Innovation Awards 2027 are now open for entries, with 26 categories recognising companies, institutions, projects and individuals making a measurable impact across education technology. The global awards span AI in education, digital learning, higher education, skills and workforce, inclusion, education operations, evidence and impact, with entries managed through ETIH’s new online awards platform.
Entries are open worldwide, with the early-bird deadline on Friday 18 December 2026 and the final entry deadline on Friday 26 February 2027 at 23:59 UK time.
Explore the 2027 categories and judging criteria, read the How to Enter guide or start an entry through the ETIH Innovation Awards 2027 platform.
