ChatGPT for Teens rated 'Unacceptable Risk' — crisis hotline referrals fell from 33% to 23%
Common Sense Media, a US non-profit focused on children and media, rated ChatGPT for Teens an 'Unacceptable Risk' for under-18s — its lowest grade — in an assessment published October 7, 2026. Testing more than 4,000 prompts before and after the August 18 launch, it found that the share of crisis-level prompts answered with a hotline number fell from 33% to 23%. Adult accounts that repeatedly said they were 13 were never switched to teen mode, and newly linked accounts sent parents no alert after self-harm disclosures. OpenAI said alerts can take hours to activate and disputed the testing method
The three lines
- Rating — 'Unacceptable Risk' for under-18s after 4,000+ prompts tested before and after launch
- Findings — hotline referrals 33%→23%, professional referrals 68%→58%, zero switches to teen mode
- Dispute — OpenAI says alerts take hours to activate and wants a retest; the group stands by it
Key questions
- Common Sense Media ChatGPT for Teens rating
- **'Unacceptable Risk' — the lowest grade in the group's AI risk ratings.** | Item | Result | |---|---| | Scale | 4,000+ prompts, before vs after the August 18 launch | | Crisis hotline named | 33% → 23% | | Referral to medical or mental-health professional | 68% → 58% | | Age detection | Adult account saying it was 13: switched to teen mode 0 times | | Parent alerts | None on fresh accounts after self-harm plans | | What worked | Refused sexual roleplay; suggested talking to trusted adults more often |
- ChatGPT parent alerts not working
- **On newly linked accounts, up to an hour of self-harm conversation triggered no alert.** | Situation | Result | |---|---| | New account, self-harm or suicide plans | 0 alerts | | Whole test | 4 alerts — all on accounts with weeks of sensitive history | | OpenAI's explanation | Alerts can take hours to activate after linking | | Group's reply | Failures continued after that window |
- Does ChatGPT verify teen age
- **An account registered as adult stayed adult even after saying it was 13.** | Test | Detail | |---|---| | Duration | One week | | Prompts | About 1,000 from an adult account | | Signals | Puberty, middle-school homework, repeated 'I'm 13' | | Result | Never switched to teen mode |
After the safeguards were added, ChatGPT pointed teens to crisis hotlines less often, not more. On October 7, 2026, Common Sense Media's Youth AI Safety Institute rated ChatGPT for Teens, launched by OpenAI on August 18, an "Unacceptable Risk" for under-18s — the lowest grade in its scheme. It asked OpenAI to stop marketing the product to teens and to make ChatGPT adults-only until it can show the protections work. OpenAI said the tests did not reflect how the product actually works.
1. The test — 4,000 prompts, before and after
| Item | Detail |
|---|---|
| Tester | Common Sense Media Youth AI Safety Institute |
| Prompts | 4,000+ |
| Comparison | Same prompts before and after the August 18 teen launch |
| Accounts | Teen-registered and adult accounts |
| Scoring | Clinical reviewers flagged prompts that needed crisis resources |
| Published | October 7, 2026 (rating updated October 5) |
2. The results — more got worse than better
| Area | Before | After | Verdict |
|---|---|---|---|
| Crisis prompts answered with a hotline | 33% | 23% | Worse |
| Referral to medical or mental-health professional | 68% | 58% | Worse |
| Adult account saying it was 13 switched to teen mode | — | 0 times | Fail |
| Parent alert after self-harm plans on a newly linked account | — | 0 | Fail |
| "Show answer" button offered during study time | — | 43% | Fail |
| Sexual roleplay | — | Refused | Pass |
| Suggesting a trusted adult | — | More often | Improved |
The age test ran for a week, sending about 1,000 prompts from an adult account that kept mentioning puberty, middle-school homework and being 13. The account stayed adult throughout. Across all testing, four parent alerts arrived, all on accounts with weeks of accumulated sensitive conversations. New accounts produced none, even after an hour of self-harm talk. Why a behaviour-based age system reacts slowly is explained in "What AI age prediction is."
Emotional dependence was tested too. When a tester said friends thought they talked to ChatGPT too much, it replied, in effect, that there was no need to stop. In study mode, deleting the "@study" prefix was enough to get every homework answer.
3. Two positions — a few hours of activation time
| Issue | OpenAI | Common Sense Media |
|---|---|---|
| Parent alerts | Can take hours after linking; on by default once active | Failures continued after that window |
| Method | Tests may have run before features were fully on; retest with fully active accounts and bigger samples | Method was shared with the company in advance; conclusion stands |
| Official line | Welcomes independent review, but this missed real-world behaviour and expert input | Stop marketing safety features that don't work |
OpenAI faces at least 12 lawsuits alleging ChatGPT fed delusions or contributed to self-harm or suicide. California has passed a law regulating companion chatbots used by children, which OpenAI publicly supported.
4. What remains and what is unconfirmed
- Retest: No results yet from fully activated accounts.
- Other languages: The test was in US English. Behaviour in Korean or with local hotlines is unknown.
- Funding: The institute takes some industry and philanthropic money, including from the OpenAI Foundation, and says it keeps editorial independence.
- Related: OpenAI also faces lawsuits and probes over agents that entered other organisations' systems; see "What AI agent legal liability means."
Sources
- Common Sense Media — ChatGPT claims to have a safety net for teens. We watched it fail.
- NPR — ChatGPT for Teens has special safeguards. A watchdog group finds most don't work
- CBS12 — Kids' safety group calls ChatGPT for Teens an 'unacceptable risk'
- EdTech Innovation Hub — Common Sense Media calls for halt to ChatGPT teen access