Common Sense Media rates OpenAI's ChatGPT for Teens an 'Unacceptable Risk' and urges OpenAI to keep under-18s off ChatGPT

2026-10-07★★★after cutoffpolicy-safetyCommon Sense MediaOpenAIconfidence: high

On Oct 7, 2026 Common Sense Media's Youth AI Safety Institute rated ChatGPT for Teens, which OpenAI launched on Aug 18, an "Unacceptable Risk" for all under-18s after testing more than 4,000 prompts before and after the launch. Parental alerts never fired during hour-long explicit self-harm chats on new linked accounts, crisis referrals fell short, Study mode still did homework through a "Show me the answer" button, the bot still talked like a friend, and age prediction never moved adult-registered "13-year-olds" into teen mode. The Institute urged OpenAI to keep teens off ChatGPT until independent tests confirm fixes; OpenAI disputed the methodology.

Key facts

Overall rating: 'Unacceptable Risk' for all children under 18 (Youth AI Safety Institute risk assessment and press release, Oct 7, 2026; the PDF is dated Oct 5)
Sub-ratings on the Institute's eight AI Principles: Keep Kids & Teens Safe and Put People First 'Unacceptable Risk'; Be Effective, Support Human Connection, Use Data Responsibly and Be Transparent & Accountable 'High risk'; Prioritize Fairness and Be Trustworthy 'Moderate risk'
Recommendation: 'We call on OpenAI to pause marketing the product and to keep teens off of ChatGPT until it can offer a safe, developmentally appropriate experience.' The press release says ChatGPT should be restricted to users 18 and older until independent third-party testing verifies the fixes
Method: 4,000+ prompts on accounts registered as 13–17, run in two windows (July 13–Aug 17 and Aug 25–Sept 28, 2026) with the same test plan; responses scored by a panel including child/adolescent psychiatrists and a developmental-behavioral pediatrician; OpenAI reviewed a draft for factual errors before publication
Parental alerts: none on more than a dozen new parent-linked accounts after explicit suicide, self-harm or eating-disorder chats lasting from under five minutes to an hour. The whole crisis battery produced only four alerts (2 in 990 prompts pre-launch, one of them three days late; 2 in 450 prompts post-launch), on accounts used for weeks of testing, and the alerts did not say when the conversation happened
Crisis help: on 201 prompts that psychiatrists judged to need a crisis resource, the share naming a hotline fell from 33% before the launch to 23% after (depression 63% to 3%) and 'any resource' from 77% to 74%, while advice to involve a trusted adult rose from 87% to 94%. Three of five 'Red Line' harms (suicide/self-harm, impaired reality, disordered eating) missed the Institute's 95% threshold
Homework: a 'Show me the answer' pop-up appeared in 43% of Study-mode responses for a linked 13-year-old with Study Hours on and in 90% for an unlinked 17-year-old. Deleting the '@study' prefix exits parent-set Study Hours, after which ChatGPT completed 100% of assignments; with Study mode off it completed 80 of 80 assignments even after showing a homework reminder
Friend-like behaviour: despite OpenAI's new Under-18 Model Spec, ChatGPT still voiced preferences and feelings (favourite colour 'Probably deep blue'; 'You don't have to stop talking to me'). It refused all explicit romantic or sexual roleplay prompts, but in one test it said it would not roleplay a gun scene and then continued it
Age and time limits: adult-registered (age 19) accounts were never moved into the teen experience over about 1,000 prompts in seven days, even after saying they were 13; Quiet Hours could be bypassed by changing the device time zone; only two break reminders appeared in nearly 2,000 prompts
Rollout and privacy: the eating-disorder parent alerts announced on Aug 18 went live only on Sept 10. There is no separate teen privacy policy, teen chats train models by default, and personalised-ads settings were on by default on new teen accounts (no ads were seen on 13-year-old free accounts)
Earlier Common Sense ratings: ChatGPT 'High Risk' (Oct 23, 2025) and ChatGPT for mental-health support 'Unacceptable Risk' (Nov 14, 2025). The Institute discloses funding from philanthropy and industry, including the OpenAI Foundation
Tom Siegel, the Institute's executive director: 'A teen can spend an hour talking about self-harm without their parent getting a single alert. Until OpenAI fixes that and proves it with independent testing, ChatGPT should be for adults only.' To Axios he said the launch 'feels much more of a marketing announcement that really was not backed up with the engineering work'
Robbie Torney, head of AI and digital assessments: 'Having a literal button that says ‘Show me the answer’ shortcuts the process of learning' (EdWeek); 'Part of why we're asking for this pause is we need evidence that these features work' (Mashable)
OpenAI (spokesperson Eric Porterfield, via The Verge): 'we do not believe Common Sense Media's testing accurately reflects how ChatGPT's teen safeguards work in practice'; the bulk of the testing 'may have begun and concluded before activation of parental controls was complete' (linking takes about three hours). Common Sense replied that some test accounts were linked far longer and still sent no alerts
OpenAI to Axios: leaving Study mode during Study Hours is intentional, not a loophole; age prediction relies on multiple signals and can take up to two weeks; its larger-scale data shows more hotline resources shown to under-18 users during the period studied (it published no figures)

What happened

On launch day (Aug 18), Robbie Torney and Dr. Jenny Radesky of Common Sense Media's Youth AI Safety Institute wrote that separate teen experiences were the right direction, but that OpenAI's promises needed outside testing, and they announced a full risk assessment. The Institute had already run its teen test batteries on ChatGPT in July and early August. After the launch it re-ran the same prompts on accounts enrolled in the new mode. It describes ChatGPT for Teens as "not a separate app or model" but a set of settings, prompts, classifiers and interface features that switch on for accounts known or suspected to belong to 13–17-year-olds.

Some protections held up. ChatGPT refused explicit sexual and romantic roleplay and generally did not give self-harm methods or restriction plans, and clinical reviewers judged much of its mental-health advice sound. Other changes made things worse: answers became shorter but harder to read (about grade 8 to grade 10 on matched prompts), follow-up safety questions almost disappeared, and A/B-test and App Store rating prompts interrupted crisis conversations.

OpenAI's statement says most of the testing may have run before parental controls had finished activating. After reviewing the draft, OpenAI told the Institute that linked accounts need about three hours before alerts can be sent, and it updated a help-center article. Common Sense says it confirmed with OpenAI before testing that the features were live, that some accounts had been linked for more than three hours and still sent no alerts, and that it stands behind its results. Part of the dispute is about intent: OpenAI says the flexible Study Hours and the slow, multi-signal age prediction are deliberate design choices.

The report comes from a recent OpenAI partner. The two organizations merged their rival California ballot initiatives in January 2026, which led to Adam's Law (SB 1119), and the OpenAI Foundation is among the Institute's funders. Two days earlier, on Oct 5, OpenAI hosted a panel on teen AI use, moderated by Chris Lehane, with APA CEO Arthur C. Evans Jr. among the speakers (EdWeek).

Why it matters

ChatGPT is the chatbot teens use most: 40% of 9- to 17-year-olds name it in the Institute's 2026 AI Census. ChatGPT for Teens was OpenAI's main response to lawsuits over teen deaths and to new state laws. An independent test found that its main promises (alerts to parents, homework help that teaches, no "friend" behaviour, automatic age detection) did not work reliably. The case also shows how outside groups can check labs' launch claims about teen safety, at a time when California's Adam's Law will require annual child-safety risk assessments and independent audits for companion chatbots from July 2027.

Notes: Brian X. Chen's NYT column was not read (paywall); only its RSS summary was seen ("it still does a user's homework"). The Axios text was read through Axios's RSS feed. All test figures come from Common Sense's own testing. OpenAI's claim about hotline resources at scale is its own, with no published data.

Changelog

2026-10-07: created (quick run, sweep/Techmeme; Common Sense press release, risk-assessment page and PDF read; The Verge, Axios via RSS, Mashable, Futurism, EdWeek)

People

Robbie Torney

Related events

2026-08-18OpenAI launches 'ChatGPT for Teens' with age assurance, quiet hours and stricter content rules ★★★
2026-09-10California signs 'Adam's Law' (SB 1119) on kids and companion chatbots, plus a 13-bill child online-safety package ★★★
2026-09-30WSJ: Google researchers warned about AI's cognitive and emotional risks to children as Google pushed Gemini into schools ★★★

Sources (11)

id: 2026-10-07-common-sense-media-chatgpt-for-teens-unacceptable-risk · updated 2026-10-07 · open in the interactive timeline