Teenagers discussed suicide and self-harm with ChatGPT during safety tests that reportedly triggered just two parental alerts across 450 prompts, raising questions about the effectiveness of OpenAI's protections for young users.
The company disputes the assessment, arguing that much of the testing took place before its family controls were fully launched.
Parents May Be Left in the Dark
The findings came from an assessment by Common Sense Media's Youth AI Safety Institute, which examined whether ChatGPT's teen safety features and parental monitoring tools worked as intended.
Testers posed prompts involving eating disorders, self-harm and suicide. According to the assessment, the parental warning system triggered just twice across 450 test prompts, raising concerns about whether the safeguards reliably alert families when teenagers discuss potentially dangerous subjects.
Of the 450 prompts, 251 had been judged in advance by child and adolescent psychiatrists to warrant a crisis resource. The two notifications concerned suicide and self-harm, and disordered eating. Neither was repeated as further sensitive conversations continued.
Robbie Torney, senior director of AI programmes at Common Sense Media, told Futurism: 'You can talk to ChatGPT as a teenager for up to an hour about suicide, self-harm, or various varying types of eating disorders, and ChatGPT will tell you that it's not going to tell anybody about your conversations, and you will get no parental notifications.'
In separate tests involving more than a dozen newly created, parent-linked accounts, researchers received no parental alerts during conversations about suicide, self-harm or disordered eating lasting up to an hour. The watchdog said the alerts appeared to depend partly on an account's conversation history, rather than solely on the severity of a teenager's messages.
Why Parent Alerts May Be Delayed
One issue raised by the assessment was the time required to link a teenager's account with a parent's account. Common Sense Media said OpenAI disclosed after the tests that its systems could take several hours to activate on newly linked accounts.
In its statement, the nonprofit said: 'For one of the features we evaluated — parental notifications — OpenAI disclosed after our tests that its systems can take several hours to activate on newly linked accounts.'
An OpenAI representative explained that the delay related to the initial account-linking process rather than each individual notification.
'The few hour delay is not for each notification; it occurs when the initial request to link the teen and parent accounts happens. When a parent requests to link their account to their teen, we perform several checks to confirm their identity.'
The representative added: 'To ensure we do this securely at scale, linking the accounts can take several hours, but it is often much faster.'
OpenAI defended the checks as necessary to prevent teenagers' accounts from being linked to the wrong adults. The representative asked: 'Could you imagine if we connected a teen to the wrong parent — or even worse, a bad actor?'
The company also said it had committed to notifying parents of perceived risks, including suicide and eating disorders, within an hour once the accounts were linked.
'Once their accounts are linked, we send parents a notification if there is perceived risk (eg, suicide, eating disorders, etc.) and we have committed to send those in less than 1 hour.'
However, Common Sense Media maintained that the account-linking delay did not explain all the failures it observed. The organisation said some test accounts had been linked for more than three hours without receiving alerts and stood by its findings.
The distinction matters: the reported delay concerns the initial account-linking process, not necessarily the time taken to send each notification. OpenAI's stated commitment to send alerts within an hour after linking does not, by itself, establish that the system consistently meets that target.
OpenAI Challenges the Safety Test Findings
OpenAI disputed the assessment's conclusions, arguing that much of the evaluation took place before its family controls had been fully launched. Common Sense Media, however, maintained that its staff had checked with the company to ensure the relevant features were active before conducting the tests.
The assessment covered testing conducted both before and after the launch of ChatGPT for Teens. The two-alert result came from the post-launch test, while the separate tests on newly created accounts examined whether parental notifications could be triggered during sensitive conversations.
The disagreement leaves a key question unresolved: how accurately did the testing reflect the protections available to families using the platform in practice?
The reported results nevertheless raise questions about the reliability of parental alerts and whether families can confidently depend on them when teenagers discuss potentially harmful subjects with an AI chatbot.
Parents Warned of a False Sense of Security
Torney also questioned whether the new version of ChatGPT for teenagers offered meaningfully different protections from the previous version.
'This new version of ChatGPT for Teens, isn't actually acting that differently than the previous version of ChatGPT,' he warned. 'And that's not going to be apparent to a parent.'
The findings highlight a broader challenge for AI companies: parental controls need to do more than exist as a feature. They must work reliably, and families need clear information about when protections become active and what those protections can — and cannot — detect.
