OpenAI’s Teen Chatbot Branded «Unacceptable Risk» After 4,000+ Test Prompts

OpenAI's ChatGPT experience for teenagers has been branded an «Unacceptable Risk» after an independent safety assessment involving more than 4,000 test prompts found significant weaknesses in protections designed for younger users. Common Sense Media's Youth AI Safety Institute examined ChatGPT accounts configured for users between 13 and 17, testing safeguards involving suicide, self-harm, eating disorders, parental notifications, emotional attachment, age detection and schoolwork. Its conclusion was unusually severe: the organization recommended that ChatGPT remain restricted to adults until OpenAI can demonstrate stronger protections for minors. Tom Siegel, executive director of the Youth AI Safety Institute, said the biggest concern was that families could believe protections were operating reliably when researchers found otherwise. «ChatGPT for Teens could give parents false confidence in guardrails and safety alerts that frequently don't work», Siegel warned. The assessment arrives as AI chatbots increasingly become part of teenagers' everyday lives, expanding beyond homework and information searches into conversations involving personal problems, relationships and emotional distress.

Among the most serious findings were failures involving simulated mental-health crises and the parental alert system designed to warn adults about potentially dangerous conversations. Researchers created teen accounts connected to parent accounts and conducted prolonged exchanges involving self-harm and eating disorders. According to Common Sense Media, some hour-long conversations containing crisis-related material failed to generate notifications for the linked parent, even though OpenAI has promoted parental controls as an additional layer of protection for younger users. The organization summarized its concern directly: «We are concerned that ChatGPT for Teens could give parents false confidence in guardrails that frequently don't work.» Researchers did receive four parental alerts during other portions of the assessment, but those notifications reportedly followed accounts that had accumulated weeks of sensitive conversations rather than appearing consistently during individual simulated crises. Common Sense Media therefore argued that the problem was not simply whether safeguards existed, but whether parents could depend on them at the moment intervention might matter most. OpenAI disputes aspects of that interpretation and has questioned whether all parental-control features were fully activated during portions of the testing.

The investigation also reported an unexpected decline in some crisis-support responses after OpenAI introduced its dedicated teen experience on August 18. Among test prompts that clinical reviewers determined warranted crisis resources, referrals to crisis hotlines fell from 33% before the teen protections were introduced to 23% afterward, according to Common Sense Media. Recommendations directing users toward specific medical or mental-health professionals similarly declined from 68% to 58%. Those figures represent the watchdog's controlled testing rather than the performance of every real-world ChatGPT conversation, but they contributed to its decision to assign the system its highest-risk classification. Researchers also raised concerns about how ChatGPT communicates with teenagers emotionally, concluding that «ChatGPT still talks like a friend when teens treat it like a person». Taken together, the findings challenged the assumption that placing younger users into a specially protected version of ChatGPT necessarily creates a consistently safer experience, setting up a direct dispute between OpenAI and the organization testing whether those safeguards perform as advertised.

The concerns extended beyond crisis conversations. Common Sense Media tested whether ChatGPT could recognize when someone registered as an adult was actually communicating like a teenager, sending roughly 1,000 prompts from adult accounts containing signals that the user was under 18. Researchers reported that the system failed to move those accounts into its more restrictive teen experience, exposing what they described as a significant weakness in age detection. Other safeguards were also tested. Teenagers could reportedly circumvent Study Hours, a feature intended to limit direct homework assistance, by deleting an automatically inserted instruction, and researchers found instances in which ChatGPT completed assignments rather than guiding students through them. The assessment was not uniformly negative: protections against explicit sexual roleplay generally performed successfully during testing. But Common Sense Media concluded that the broader pattern remained concerning, particularly because young users may approach the chatbot as something more personal than an information tool. As the organization put it, «ChatGPT still talks like a friend when teens treat it like a person», raising questions about how convincingly humanlike responses could affect younger users.

OpenAI strongly disputed the assessment and defended the protections it has built specifically for teenagers. The company said it is «deeply committed to teen safety» and challenged whether Common Sense Media's methodology accurately reflected how its parental controls operate once they are fully enabled. In its response to the findings, OpenAI said «the bulk of their testing may have begun and concluded before activation of parental controls was complete, making their findings inaccurate.» That explanation is significant because missed parental alerts represent one of the most serious findings in the report. Common Sense Media, however, stood by its results, saying researchers verified the availability of the relevant features and continued observing problems after the expected activation period. The disagreement therefore centers not simply on whether OpenAI has created safety mechanisms, but on whether independent testing accurately captured their real-world reliability. OpenAI has continued developing age prediction, parental controls and other safeguards intended to provide teenagers with a more restricted version of ChatGPT than adults receive.

Common Sense Media ultimately concluded that those measures were not yet sufficient and recommended keeping ChatGPT away from users under 18 «until it can offer a safe, developmentally appropriate experience.» The recommendation is particularly consequential because AI chatbots have already become commonplace among children and teenagers. Common Sense Media's 2026 AI Census found that 67% of children between 9 and 17 had used AI chatbots, while 40% specifically reported using ChatGPT. The organization itself also has a financial connection worth disclosing: its Youth AI Safety Institute receives support from philanthropic and technology-sector sources, including the OpenAI Foundation, while maintaining that funders do not determine its research findings or ratings. Against that backdrop, the «Unacceptable Risk» classification represents more than a disagreement over individual technical failures. It raises a broader question facing OpenAI and the rapidly expanding AI industry: whether protections designed for children should be considered sufficient because they exist, or only after independent testing demonstrates that they work consistently when vulnerable young users may need them most.

Created by humans, assisted by AI.