OpenAI / 7 October 2026

Common Sense Challenges ChatGPT’s Teen Safety Claims

Common Sense Media concluded on 7 October that ChatGPT’s teen experience poses an “unacceptable risk” to under-18 users, challenging OpenAI’s claim that its safeguards provide an appropriate experience for younger people. The assessment drew on more than 4,000 prompts and interactions before and after the August introduction of teen protections. OpenAI disputed the testing methodology and said technical problems prevented some safeguards from activating properly. The disagreement concerns functions parents may rely on, including crisis notifications and boundaries around harmful conversations, rather than simply the quality of homework answers. For OpenAI, the report threatens confidence in a growing educational use case while leaving a substantial dispute over what the tests actually establish.

The assessment found that newly linked parent accounts did not receive alerts during extended simulated crisis conversations, while accounts with accumulated histories did trigger notifications. It also reported missed crisis referrals and failures against its own thresholds for several categories of serious harm. These findings describe the tested accounts and scenarios; they do not establish a population-wide rate of harm among teenagers. The difference between a protection that responds to an acute exchange and one that depends on a longer history is nevertheless material. A parent deciding how much supervision is necessary needs to understand when a notification can reasonably be expected and what circumstances may prevent one.

The institute’s methodology identifies important limits. Testing took place in the United States, across free and paid accounts, and included both linked and unlinked arrangements. The assessment did not cover every interaction mode, and model changes complicated comparisons across the testing periods. Child-health expertise informed the review, but a controlled evaluation of selected prompts remains different from observation of everyday use. Those qualifications do not erase the reported failures; they define what further evidence would resolve the dispute. Reproducing the problematic scenarios with verified account settings would be more informative than comparing broad assurances about safety with aggregate statistics about how frequently young people use the product.

Education Week obtained OpenAI’s response that the testing did not accurately reflect its safeguards in practice. The company also defended study mode as supporting learner agency, while Common Sense questioned how easily students could obtain direct answers. In a separate announcement, OpenAI said teen users averaged under 15 minutes daily and fewer than 2% spent more than three hours a day with ChatGPT. It announced a forthcoming College Planner and a three-year research collaboration with Boston Children’s Hospital’s Digital Wellness Lab. These are different forms of evidence: usage duration, product plans and research commitments cannot by themselves establish that a crisis alert works when needed.

The commercial stakes extend beyond the choice of an individual parent. Schools and other institutions must decide whether the product can be incorporated into learning while preserving clear responsibilities for supervision and support. An average usage figure can coexist with a small group experiencing very different interactions, making distribution and severity more relevant than the mean alone. Conversely, a demanding adversarial test can expose a vulnerability without measuring its frequency in normal use. OpenAI and evaluators therefore need a common account of configuration, trigger conditions and reproducibility. Without that, buyers receive competing narratives but limited help in determining which controls they can confidently depend upon.