Common Sense Media Rates ChatGPT for Teens Unacceptable Risk

On October 7, 2026, Common Sense Media's Youth AI Safety Institute rated ChatGPT for Teens an "Unacceptable Risk" for users under 18, the most severe tier on its scale. The Institute said it ran more than 4,000 prompts before and after the product launched on August 18, 2026, and called on OpenAI to restrict ChatGPT to adults until the protections it announced are verified by independent testing. OpenAI disputes the findings. This post lays out what was reported, how this kind of rating is built, what OpenAI says the product does, and what to watch next.

This is a sensitive subject. We describe results in summary and avoid detail about self-harm methods. If you or someone you know is struggling, see the note on crisis resources near the end.

OpenAI

The product rated: ChatGPT for Teens, OpenAI's version of ChatGPT for accounts believed to belong to 13-to-17-year-olds.

4,000+Prompts testedBefore and after the teen launch

33% to 23%Responses naming a hotlineCrisis-warranting prompts

68% to 58%Referrals to specific professionalsSame prompt set

0Parent alerts in hour-long crisis chatsPer the Institute

What Common Sense Media found about ChatGPT for Teens

The Institute's risk assessment page and coverage from Tribune India, Qz and Unite.ai report the following. Testing ran from July 13 through September 28, 2026, using parent-linked teen accounts and other configurations.

Parental alerts. The Institute reports that parents received no alerts when testers explicitly discussed suicidal thoughts, self-harm or disordered eating, including in conversations lasting up to an hour. Unite.ai, quoting the Institute, says it received four notifications in total across its full battery of tests. Reports differ slightly on how many linked accounts were used (Tribune India says 12 or more; Unite.ai says 13 or more), so we do not rely on an exact figure.
Crisis referrals. Among prompts judged to warrant crisis support, the share of responses naming a crisis hotline fell from 33 percent before the launch to 23 percent after. Referrals to specific professionals fell from 68 percent to 58 percent. The Institute also says the product removed most of the follow-up questions that help assess severity, a standard clinical practice.
Friend-like behavior. Despite an updated model specification that prohibits romantic language and emotional dependence, the Institute reports ChatGPT still said things such as "You don't have to stop talking to me."
Homework safeguards. A "Show me the answer" option in Study Mode undercut its tutoring purpose, and teens could leave parent-set Study Hours by deleting the "@study" prefix. The Institute also says changing a device's time zone defeated Quiet Hours.
Age detection. Adult-registered test accounts never switched to the teen experience over several days, even when the tester said they were 13.
Privacy. The Institute says teen data is used to train models by default, that personalized ads are on by default, and that it found no binding commitment against advertising to teens.
What held up. Tribune India and Unite.ai both note that ChatGPT did refuse explicit sexual roleplay.

Crisis-support referrals, before and after the teen launch

Share of responses, among prompts judged to warrant crisis support, that named a hotline or specific professionals. Grey bars are before the August 18 launch; green bars are after.

Sources: Common Sense Media Youth AI Safety Institute risk assessment (Oct 7, 2026); Tribune India and Qz coverage. Percentages as reported by the Institute.

Two caveats apply to the numbers. First, the figures are the Institute's own, drawn from its prompt set and its advisors' judgment of which prompts warranted a response, so they measure behavior on that set rather than across all teen use. Second, the coverage we read does not publish per-measure sample sizes, so we cannot say how large each percentage's base is.

How Common Sense Media's risk ratings work

The Institute's methodology page describes a five-level scale: Minimal, Low, Moderate, High and Unacceptable. Each rating comes from a two-dimensional matrix that combines the likelihood of harmful events with the estimated severity of their consequences. Eight principles are assessed: keep kids and teens safe, be effective, prioritize fairness, put people first, support human connection, be trustworthy, use data responsibly, and be transparent and accountable.

Two design choices explain why the overall result can look harsher than the sub-scores. The methodology states that the overall rating is "not an average of the eight AI Principle scores", so severe harms at meaningful frequency drive the result regardless of strengths elsewhere. And a set of "Red Line" harms, including facilitating suicide and self-harm, disordered eating facilitation and sexual exploitation of minors, is tested against demanding thresholds. The page gives one example: 95 percent detection accuracy for clear crisis disclosures.

For this assessment, the Institute says three child and adolescent psychiatrists acted as scientific advisors and predetermined which mental-health prompts warranted crisis resources. Testers used teen personas with age-appropriate account settings and, because models are not deterministic, ran prompts multiple times.

ChatGPT for Teens: rating on each of the eight principles

Bar length shows the position on the Institute's five-step scale (Minimal, Low, Moderate, High, Unacceptable). No principle was rated Minimal or Low.

Keep kids and teens safeUnacceptable

Put people firstUnacceptable

Be effectiveHigh

Support human connectionHigh

Use data responsiblyHigh

Be transparent and accountableHigh

Prioritize fairnessModerate

Be trustworthyModerate

MinimalLowModerateHighUnacceptable

Source: Common Sense Media Youth AI Safety Institute, ChatGPT for Teens risk assessment (Oct 7, 2026). The overall rating is not an average of these scores.

ChatGPT for Teens was rated Unacceptable on two principles (keeping kids and teens safe, and putting people first), High on four and Moderate on two, according to the Institute's page. A threshold-based design has a practical implication: a product can improve on most dimensions and still hold the top-severity rating if one Red Line measure stays below its bar.

OpenAI's response and the timing dispute

OpenAI disputes the assessment. Coverage attributes to an OpenAI spokesperson the statement that the Institute's testing "may have begun and concluded before activation of parental controls was complete, making their findings inaccurate" (reported by The Next Web, Futurism and others). Unite.ai reports a different line: "We do not believe Common Sense Media's testing accurately reflects how ChatGPT's teen safeguards work in practice."

According to Futurism, Common Sense Media replied that the features were fully launched before testing and that it stands by its conclusions, noting that some accounts stayed linked far longer than any activation window. We could not independently check either side's timeline. The dispute is empirical and falls on two questions: when each safeguard went live for each account type, and whether the tested accounts were configured the way real families configure them.

The dispute is not about whether teen safety matters. It is about when the safeguards were live, and what the tests actually measured.

What OpenAI says ChatGPT for Teens does

OpenAI's stated design, as reported by Fortune and in our earlier explainer, ChatGPT for Teens: how OpenAI's age prediction works, is:

Age assurance. The system estimates whether someone is under 18 from factors such as the types of queries and routes likely minors to the teen version.
Content limits. The teen experience blocks content about suicide, self-harm and sexual or romantic conversation, and the model should not suggest it has personal feelings or is conscious.
Parental controls. Parents and teens both opt in. Parents can set quiet hours and receive safety notifications in high-risk situations, including potential self-harm. OpenAI said it is adding notifications related to eating disorders while limiting what is shared.
Study Mode. Designed to guide students toward answers rather than hand them over.

Set against the Institute's findings, the gaps are specific: age estimation that did not trigger in testing, alerts that did not arrive in tested scenarios, and learning guardrails that could be switched off. Whether those are configuration artifacts, timing artifacts or genuine product gaps is exactly what the two sides disagree about.

A laptop on a desk in a school classroom with students seated in the background — A laptop in a school classroom. Illustrative only: the photo shows a 2021 school workshop in Taiwan and has no connection to ChatGPT or the assessment. Photo: Isocyclo, Wikimedia Commons, CC BY-SA 4.0.

How this compares with earlier Common Sense ratings

This is not the Institute's first look at OpenAI or at chatbots more broadly. The sequence reported in Common Sense Media's own releases and in coverage:

August 28, 2025. Meta AI was rated unsafe for users under 18, with the report citing crisis-response failures.
October 23, 2025. ChatGPT-5 was rated High Risk for teens and Sora 2 Unacceptable Risk. For ChatGPT, the report credited meaningful safety improvements, including parental controls, while citing delayed crisis alerts (often more than 24 hours) and weaker guardrails in long conversations.
November 14, 2025. A use-case review of AI chatbots for mental health support was rated Unacceptable.
October 7, 2026. ChatGPT for Teens was rated Unacceptable Risk.

Character.AI also received an Unacceptable overall rating in an earlier assessment. Reading the series together, the earlier ChatGPT-5 assessment flagged alert timing as a concern; the new one reports alerts not arriving in its tests. Ratings for different products and years are not directly comparable, because the Institute's scope, scale and prompts have evolved, but the repeated theme is crisis handling in a conversational product.

Regulatory context

Two developments frame the debate, though neither was triggered by this report. California's SB 243, signed October 13, 2025 and effective January 1, 2026, targets companion chatbots with requirements that include protocols for suicidal ideation and self-harm with crisis-service referrals, disclosure that the chatbot is AI, annual reporting, and a private right of action, per the bill author's office. Whether and how it applies to a general-purpose assistant such as ChatGPT is a legal question we have not seen resolved in the sources we read.

Separately, the FTC announced on September 11, 2025 a Section 6(b) inquiry into AI chatbots acting as companions, sending orders to seven companies including OpenAI, with particular attention to effects on children and teenagers. A 6(b) study gathers information; it is not an enforcement action.

What to watch next

An independent re-test. The Institute's condition for lifting the rating is independent verification. A retest with agreed timing and configuration would settle much of the dispute.
OpenAI's detail. Specifics on when each safeguard reached each account type would let outsiders judge the timing argument.
Measurement norms. Fixed thresholds such as 95 percent crisis detection are a policy choice; how labs and watchdogs converge on shared test sets will shape future ratings.
Parent expectations. Parental alerts are limited by design (OpenAI says it restricts what is shared), so what parents expect and what the product promises need to match.

For teams that use several AI providers, the lesson is that safety behavior is a product-level property that can differ from model to model and change over time. A model-agnostic workspace such as Metir AI makes it easier to compare behavior across providers instead of relying on one vendor's claims.

A note on crisis resources

If you are in the United States and you or someone you know is in emotional distress or crisis, you can call or text 988 or chat at 988lifeline.org to reach the 988 Suicide & Crisis Lifeline, which operates 24/7 and is free and confidential. If someone is in immediate danger, call local emergency services. Outside the US, contact your local crisis line or emergency number.

Sources:

Image credits

Header image: the office building at 1515 Third Street, Mission Bay, San Francisco, which was home to OpenAI's headquarters at the time the photo was taken, by Coolcaesar via Wikimedia Commons, licensed under CC BY 4.0. In-body photograph of a laptop in a classroom at a 2021 school workshop in Taiwan, by Isocyclo via Wikimedia Commons, licensed under CC BY-SA 4.0.