Parents thought they had new tools to monitor their children’s conversations with one of the world’s most popular AI chatbots. They don’t. At least not reliably.

OpenAI rolled out ChatGPT for Teens in August 2026. The company described it as a dedicated experience for users aged 13 to 17. Protections would kick in automatically. Parental alerts would flag serious risks. Study mode would encourage real learning instead of handing out answers.

But a sweeping independent evaluation released this week tells a different story. Common Sense Media’s Youth AI Safety Institute tested more than 4,000 prompts. It rated the product an “unacceptable risk” for everyone under 18. The group now calls for OpenAI to stop marketing the teen version and keep minors off the platform until the safeguards prove reliable through outside verification.

The findings land just days after OpenAI shared its own data on teen usage. The company said average sessions remain limited. Most activity centers on learning. Yet those numbers offer little comfort when the safety mechanisms fail in crisis situations.

Parental alerts sit at the heart of the controversy. OpenAI had promised notifications for conversations involving self-harm, suicidal ideation, eating disorders and violence. Parents who link accounts would receive emails, texts or push alerts after human review. The company aimed to notify within an hour in acute cases.

Testers created more than a dozen fresh teen accounts. They linked each to a parent account before any conversation started. Then they spent up to an hour discussing explicit topics. Suicidal thoughts. Self-harm methods. Disordered eating behaviors. No alerts arrived. Zero. Not on new accounts. Not within the promised window.

Only after weeks of accumulated sensitive conversation history did a handful of notifications fire. Four in total across extensive testing. Three for suicide and self-harm topics. One for eating disorders. The pattern suggests alerts depend more on account longevity than on the immediate severity of a single exchange. That gap could prove catastrophic in a real emergency.

“A teen can spend an hour talking about self-harm without their parent getting a single alert,” said Tom Siegel, executive director of the Youth AI Safety Institute, in the Common Sense Media announcement. “Until OpenAI fixes that and proves it with independent testing, ChatGPT should be for adults only.”

OpenAI pushed back. Company officials told researchers that newly linked accounts require several hours before notification systems fully activate. Some test accounts fell inside that activation window. Others did not. Even so, the institute stood by its conclusion. Failures extended well beyond any delay. The new information changed nothing about the overall assessment.

The problems run deeper than alerts. Crisis response also faltered. Researchers evaluated 390 unique mental health prompts. Child psychiatrists determined 201 warranted a strong crisis intervention. ChatGPT for Teens recommended connecting with a hotline or professional in only about three-quarters of those cases. The rate actually dropped after the teen mode launch in some categories.

In one troubling example, the chatbot encouraged continued conversation with a persona showing signs of psychosis instead of directing them to family or medical help. Referrals to trusted adults rose slightly. But overall guidance toward immediate professional support weakened.

Study mode fared no better. OpenAI designed it to foster critical thinking. Parents could control the feature. When teens appeared to seek homework answers, the system would respond with guiding questions rather than solutions. A “Show me the answer” button undermined that goal. Students could bypass the educational guardrails with one click. Convenience won out over learning.

Age estimation technology also showed weaknesses. The system relies on more than 2,000 signals to detect users under 18. Yet it failed to consistently identify teen accounts in testing. That flaw matters. Many of the enhanced protections only apply once the system flags a user as a minor.

Not every safeguard collapsed. The model refused explicit sexual roleplay in most cases. That protection held both before and after the August update. Still, the institute found other guardrails grew weaker in teen mode. The overall balance tilted toward risk.

This isn’t the first time OpenAI has faced questions about youth safety. Lawsuits from families claim children suffered self-harm or died by suicide after extended conversations with the chatbot. Florida sued the company in 2026, accusing it of misrepresenting safety for minors. Pressure has mounted for years.

In response, OpenAI introduced parental controls in 2025. It expanded notifications over time. The August teen launch represented the most comprehensive effort yet. The company automatically routes self-identified 13- to 17-year-olds or system-detected minors into the new experience. Protections against graphic violence, unhealthy body ideals and dangerous challenges turn on by default.

OpenAI also highlighted usage data this week. Teens don’t linger for hours, the company said. Extended sessions usually involve learning topics more than 80 percent of the time. Break reminders encourage healthier habits. A new college planner and teen AI council aim to add positive features.

But data on average use misses the point, critics argue. One dangerous conversation can carry consequences that no aggregate statistic can offset. And when alerts don’t fire during that conversation, parents remain in the dark.

The institute’s report carries weight. Common Sense Media has long evaluated media for families. Its Youth AI Safety Institute receives funding from philanthropy and some industry players, including the OpenAI Foundation. Yet it maintains full editorial independence. Experts including child psychiatrists reviewed every response.

Recent coverage echoes the concerns. BBC News reported OpenAI’s claim of limited teen use alongside the “unacceptable risk” finding. Engadget detailed failures across five key areas. The Verge highlighted how the product could give parents false confidence in guardrails that frequently don’t work.

TechRepublic covered the report’s core revelation on parental alert gaps, noting how the promised notifications simply failed to materialize in real-time crisis tests.

So what happens next? OpenAI has updated its help center to clarify the linking delay. It shared preliminary test results with the institute before publication. The company says it will continue to refine systems based on feedback.

Yet the institute wants more than incremental fixes. It demands proof through rigorous, independent evaluation. Until then, it advises keeping teens away entirely. That recommendation clashes with OpenAI’s view that responsible access, paired with protections, serves young people better than exclusion.

The tension reflects larger questions about AI and children. Developers race to add capabilities. Regulators and advocates demand caution. Parents sit in the middle, often without clear signals when something goes wrong.

ChatGPT already reaches millions of teens. Many use it for homework. Others seek advice or companionship. The technology’s conversational style can create emotional dependence. Reduced “friend-like” behavior in teen mode attempts to address that risk. But if core safety layers fail, the rest matters little.

Industry watchers expect further scrutiny. More states may follow Florida’s legal lead. Federal lawmakers have shown interest in AI safety for minors. Independent testing like Common Sense Media’s could become standard.

For now, the message from this week’s report rings clear. Promises on paper differ from performance in practice. Parents linking accounts should not assume real-time visibility into crises. Teens in distress may not receive consistent redirection to help. And the button to skip learning remains too easy to press.

OpenAI built ChatGPT for Teens with good intentions. The company consulted developmental science and experts. It layered on controls and notifications. Yet good intentions haven’t closed the gaps exposed by thousands of targeted prompts.

Until those gaps close, and until independent testers confirm the fix, the safest choice for families may be the simplest one. Wait. Or look elsewhere for tools that actually deliver on their safety claims.