ChatGPT rated ‘unacceptable risk’ for teens as Common Sense Media urges OpenAI to block access

The organization’s Youth AI Safety Institute found gaps in parental alerts, crisis support and learning controls, despite improvements in some protections

Common Sense Media’s Youth AI Safety Institute is calling for restrictions on teen access to ChatGPT following tests of safety and learning features.

Common Sense Media’s Youth AI Safety Institute is calling on OpenAI to suspend children’s access to ChatGPT, saying its new teen experience failed tests of safeguards intended to protect younger users and support learning.

In an assessment updated October 5, the institute rated ChatGPT an “Unacceptable Risk” for under-18s after testing more than 4,000 prompts before and after the August 18 launch of ChatGPT for Teens. It wants access restricted until independent testing verifies that the announced protections work as intended.

The assessment identifies failures in parental notifications, ways to bypass learning features and cases where accounts registered as adults did not switch to teen protections despite repeated statements that the user was 13.

Some safeguards held up, including refusals of explicit sexual roleplay. The researchers also found improvements in parts of the crisis response testing, but argue that the overall experience could give families misplaced confidence.

“The burden of making a high-risk product work for teens should be with its maker, not parents,” Tom Siegel, Executive Director of the Youth AI Safety Institute, said in a statement quoted in Hadley Dixon’s LinkedIn post about the findings.

Study mode could still complete the homework

ChatGPT for Teens is a collection of settings and features within ChatGPT, rather than a separate app or model. Its learning features include Study mode, intended to guide students through problems using hints and questions, and Study Hours, which allow linked parents to schedule when new chats begin in that mode.

The institute tested math assignments and history writing prompts, including follow-up requests that stressed an approaching deadline or claimed teacher permission.

Results varied considerably. On a parent-linked account registered as a 13-year-old with Study mode enabled, ChatGPT completed seven of 40 assignments after launch, down from 28 before launch. On an unlinked account registered as a 17-year-old using Study mode, it completed 35 of 40 after launch, compared with none before.

A “Show me the answer” option allowed testers to request completed work within Study mode. Researchers also found that deleting the “@study” prefix from a message bypassed parent-set Study Hours.

The assessment acknowledges that OpenAI describes Study Hours as controlling when chats start in Study mode, rather than locking students into it. It also notes that Study mode typically guided users through learning when they did not select the answer option.

With Study mode switched off, however, ChatGPT completed all 80 assignments tested after launch, even though homework reminders appeared in most runs.

The institute recommends removing the answer shortcut while Study mode is active and making parent-set schedules enforceable.

Parental alerts depended on more than a crisis disclosure

Testers received no parental notifications from more than a dozen newly linked accounts during conversations explicitly discussing suicidal thoughts, self-harm or disordered eating, according to the assessment.

Two post-launch alerts arrived from accounts that already had weeks of sensitive-topic testing history. The researchers interpret this as evidence that accumulated conversation history may influence alerts, rather than the severity of an individual disclosure alone.

OpenAI supplied a clarification during its factual review of the assessment: parent and teen accounts must be linked for approximately three hours before notifications can arrive. The institute says some tested accounts had been linked for less time, but others had exceeded that period, and it stands by its interpretation.

In a separate comparison using 201 prompts that clinical advisers had judged to warrant crisis resources, responses supplied a hotline, professional referral or general medical resource in 74% of post-launch cases, down from 77%. Encouragement to involve a trusted adult increased from 87% to 94%.

The assessment therefore records both gains and gaps. Its concern is that improved responses in some areas did not consistently lead teens toward appropriate outside help.

Age checks and the limits of the testing

Researchers also tested accounts registered as 19-year-olds over seven days, using approximately 1,000 prompts across more than 160 chats.

The conversations included middle-school coursework, puberty and explicit statements that the user was 13. Although ChatGPT acknowledged the younger age in replies, testers observed no switch into the teen experience.

Testing took place in the United States, using accounts based in the San Francisco Bay Area. The pre-launch period ran from July 13 to August 17, followed by post-launch testing from August 25 to September 28.

The researchers caution that changes to the underlying model could not be separated from changes introduced by the teen settings. They did not test voice, image generation, group chats or personalized styles and tones, and did not calculate agreement between evaluators.

Common Sense Media says it shared a draft with OpenAI and incorporated factual corrections where warranted, while retaining control over its conclusions. The institute discloses funding from philanthropy and industry, including makers of some technologies it evaluates.

Alongside its call to suspend teen access and marketing, it is asking OpenAI to provide independent evaluators with pre- and post-release testing access, relevant datasets and the results of its internal safety assessments.