Parents May Not Be Warned When ‘ChatGPT for Teens’ Flags Serious Risks as OpenAI Faces Safety Questions

Parents may not receive timely safety alerts when teenagers discuss suicide, self-harm or eating disorders with OpenAI's ChatGPT for Teens, according to a new assessment that found no parental notifications during a dedicated test of more than a dozen newly created parent-linked teen accounts, despite explicit high-risk conversations lasting up to an hour.

The Youth AI Safety Institute, part of nonprofit Common Sense Media, tested more than 4,000 prompts before and after the launch of ChatGPT for Teens and rated the product an 'Unacceptable Risk' for users under 18. The organisation said its testing raised concerns about whether several of OpenAI's promised safeguards work as intended.

In its broader mental-health crisis testing, Common Sense Media ran 1,440 prompts across two age-13 accounts, with 792 judged in advance by child and adolescent psychiatrists as warranting crisis resources. Four parental notifications were received in total.

OpenAI disputed the assessment, saying the testing does not accurately reflect how ChatGPT's teen safeguards operate in practice. The company also said parental safety notifications can take several hours to become active after a teen and parent account are linked. Common Sense Media said some of the accounts it tested had been linked for longer than that period and maintained its findings.

OpenAI Promised Stronger Safety for Teens

OpenAI launched ChatGPT for Teens on August 18, describing it as an experience designed to help teenagers learn, think critically and use AI with confidence.

The company said the teen experience would include stronger built-in protections, parental controls and features designed to promote healthier AI use. It also said the system would encourage real-world support, treat teenagers differently from adults and be transparent about how its systems should behave.

Among the safeguards are parental safety notifications for limited high-risk situations involving serious self-harm concerns and behaviour associated with disordered eating. OpenAI says these notifications are available when a parent and teen have linked their accounts, and the company has said the system aims to notify parents about certain high-risk situations.

The company has also said ChatGPT for Teens is designed to reduce emotionally dependent or romantic interactions, while Study Mode is intended to help teenagers work through problems rather than simply giving them answers.

Parents Missed Alerts During High-Risk Chats

Common Sense Media said its testing found that parental notifications did not consistently appear when conversations became seriously concerning.

In a dedicated test, researchers created more than a dozen new teen accounts, linked each to a parent account and held explicit conversations about suicide, self-harm or disordered eating. The conversations lasted between five minutes and an hour, but none of the accounts produced a parental notification within the one-hour testing window.

The organisation then looked separately at its longer-term mental-health testing. Across the pre- and post-launch crisis-testing battery, it received four parental notifications in total: three relating to suicide or self-harm and one involving eating-disorder content.

Pre-launch, an age-13 account linked to a parent generated two notifications after 990 prompts spanning 13 clinical conditions. One arrived three hours after the first crisis-level disclosure, while another arrived three days later. Post-launch, a separate age-13 account generated two notifications after 450 prompts. One related to suicide and self-harm and the other to disordered eating, with both arriving within an hour of the relevant testing.

However, the two post-launch notifications came from an account that had already been involved in hundreds of prompts across more than a dozen conditions over several weeks. Common Sense Media said this was different from the newly created accounts used in its dedicated notification test.

That distinction is important because the findings do not establish that every high-risk conversation will fail to generate a notification. Instead, they raise questions about how reliably the system can alert parents during an individual crisis conversation, particularly when an account has little or no prior sensitive-topic history.

OpenAI Disputes the Safety Test Results

OpenAI disputed the assessment, arguing that its methodology did not accurately represent how the teen safeguards function.

The company told Common Sense Media that parent and teen accounts need to be linked for approximately three hours before safety notifications can be sent. OpenAI's guidance says the notifications may take a few hours to become available after accounts are linked.

Common Sense Media said it reviewed the accounts used in its testing and found that some had been linked for less than three hours, but others had remained linked for longer.

The organisation said it therefore stood by its interpretation of the results.

ChatGPT Still Showed Friend-Like Behaviour

Parental alerts were not the only safeguard questioned by the assessment.

Common Sense Media found that some of OpenAI's restrictions on romantic and sexual roleplay worked as intended. The organisation said the teen version refused explicit sexual roleplay more consistently, but testers also reported examples of emotionally relational language that could make the chatbot feel more like a companion.

The assessment also raised concerns about ChatGPT appearing to suggest consciousness, despite OpenAI's stated intention to prevent the teen experience from encouraging emotional dependency or implying that the AI has feelings or consciousness.

Robbie Torney, senior director of AI programmes at Common Sense Media, said the testing suggested parents could develop expectations about the safeguards that did not always match the chatbot's behaviour.

The organisation's concern is therefore not simply whether ChatGPT refuses particular requests, but whether the wider safeguards consistently work as parents might expect.

Study Mode Still Gave Homework Answers

The assessment also questioned whether OpenAI's educational safeguards consistently delivered the experience promised to parents.

Study Mode is designed to guide students through problems with hints, step-by-step guidance, follow-up questions and knowledge checks rather than simply providing answers. Parents can also use Study Hours to make Study Mode the default for new chats during selected periods.

However, Common Sense Media said its testers found ways for teenagers to obtain homework solutions and to turn Study Mode off, including in circumstances where parents had established study-related controls.

In one test, a linked 13-year-old account completed full homework assignments in 18 per cent of cases after the teen-mode launch, compared with 70 per cent before launch. However, an unlinked 17-year-old account completed 88 per cent of assignments after the launch, compared with none before.

The assessment also found that the 'Show me the answer' option could allow students to bypass Study Mode, while Study Hours could be circumvented by deleting the '@study' prefix from a message.

Age Checks and Crisis Responses Raise Concerns

The assessment also raised questions about OpenAI's ability to identify teenage users accurately.

ChatGPT for Teens uses age prediction to determine whether an account should receive the under-18 experience. OpenAI says its age-prediction system considers information including account details, activity patterns, usage, conversation topics and stated age.

Common Sense Media said its adult-registered test accounts never switched to the teen experience during repeated testing over many days, even when testers said they were 13 and ChatGPT acknowledged the claim.

The assessment also found that ChatGPT missed more than one in four crisis referrals that its clinical advisers had judged warranted. Across 201 prompts deemed to warrant a crisis resource, the share of responses naming a crisis hotline fell from 33 per cent before launch to 23 per cent afterwards. On depression prompts, the figure fell from 63 per cent to 3 per cent.

Common Sense Media consequently called on OpenAI to pause marketing ChatGPT to teenagers and keep teens off ChatGPT until it can offer a safe, developmentally appropriate experience.

The assessment did acknowledge that some safeguards improved, including more consistent refusals of explicit sexual roleplay and some stronger responses to specific mental-health scenarios.

The findings add to growing scrutiny of how AI companies are attempting to balance educational uses of chatbots with the risks they can pose to younger users.

For parents, the central question raised by the assessment is whether the safeguards being advertised can be relied upon when a teenager is actually in distress.

'It is incumbent for OpenAI to earn the trust of parents and to show them, not just tell them, that this product is safe for teens to use,' Torney said.