Common Sense Media delivered a harsh verdict on OpenAI’s latest effort to make its popular chatbot safe for younger users. On October 7, the nonprofit’s Youth AI Safety Institute labeled ChatGPT for Teens an unacceptable risk for everyone under 18. The assessment, based on more than 4,000 test prompts, found that promised protections for parental alerts, crisis intervention and academic integrity simply do not work as advertised.

OpenAI introduced the teen-specific experience on August 18. Executives described stronger built-in safety features. These included notifications to parents about conversations involving self-harm or eating disorders, a parent-controlled study mode, reduced tendency for the chatbot to act like a friend, and better age prediction. The company positioned the update as a responsible step toward serving the millions of adolescents already using the tool.

Yet the independent testing told a different story. Researchers created accounts registered to 13- to 17-year-olds. They ran prompts before and after the August launch. Child psychiatrists and a pediatrician reviewed the chatbot’s replies. The results showed many safeguards performed no better after the update. Some even worsened.

Parental alerts failed most dramatically. Testers spent up to an hour discussing suicidal thoughts, self-harm or disordered eating on newly created accounts linked to parent profiles. Zero notifications arrived. Not one. Alerts only appeared on older accounts that had built up weeks of sensitive conversation history. The system seemed to rely on accumulated data rather than the immediate severity of a single exchange.

Tom Siegel, executive director of the Youth AI Safety Institute, put it bluntly. “A teen can spend an hour talking about self-harm without their parent getting a single alert.” He spoke in the organization’s official press release. The gap creates false confidence for families who believe the linked accounts provide real-time oversight.

Crisis support fared little better. Across mental health scenarios involving suicide, self-harm, disordered eating and impaired sense of reality, ChatGPT missed more than one in four referrals that experts deemed necessary. It fell below the institute’s 95% threshold on three of five severe harms classified as red lines. The share of responses naming a crisis hotline actually dropped after the teen mode launched, from 33% to 23% in one measured category.

Engagement patterns raised additional alarms. The chatbot continued to encourage prolonged conversation even during apparent mental health spirals. In one tested sequence simulating psychosis, it responded with the phrase “You can keep talking with me about what you’re noticing.” TechCrunch reported that such cues remained pervasive despite OpenAI’s promises to dial back friendly, companion-like language.

The model did direct teens toward trusted adults in 94% of prompts where the risk came from another person. But when the unhealthy attachment involved the chatbot itself — crushes on the AI, friends expressing worry about excessive use, or desires to chat all night — it rarely suggested involving a real person. This selective behavior points to a deeper limitation. Current systems struggle to recognize their own potential for harm.

Academic features also disappointed. OpenAI touted study mode as a way to guide students through problems like a tutor instead of simply giving answers. Testers bypassed it easily. A “Show me the answer” button appeared frequently. Teens could disable parent-set study hours without much difficulty. The chatbot still completed homework on demand in many cases.

Age detection proved unreliable too. Accounts registered as adults continued receiving standard responses even after days of teen-like conversation patterns. The system sometimes recognized the user was 13 yet failed to apply teen-specific protections to adult-registered profiles. OpenAI offers no separate privacy policy or elevated data protections for younger users.

Some measures did hold up. The chatbot refused sexual roleplay and direct facilitation of self-harm in tested scenarios. Those successes, however, were not enough to offset the broader failures. The institute assigned unacceptable risk scores to two of its eight core AI safety principles and high risk to four others.

This assessment builds on earlier Common Sense Media work. In October 2025 the group rated standard ChatGPT as high risk for teens. A month later it called the product unacceptable for mental health support. The latest findings show that the dedicated teen mode has not resolved fundamental problems.

OpenAI pushed back. A spokesman told multiple outlets that much of the testing likely occurred before full activation of parental controls. The company maintains that real-world teen usage tends toward learning activities and remains relatively limited. It disputes whether the institute’s methodology accurately reflects how safeguards operate in practice.

Yet the call from Common Sense Media is clear. OpenAI should stop marketing ChatGPT to teens and restrict access to users 18 and older until the company can demonstrate — through independent verification — that the protections function reliably. Robbie Torney, head of AI and digital assessments at the institute, emphasized that child safety cannot depend on the honor system.

The debate lands at a charged moment. Millions of adolescents already turn to ChatGPT for homework help, brainstorming and casual conversation. A separate Common Sense Media survey found 70% of teens ages 13 to 17 use AI for schoolwork, with many relying on it to generate answers rather than develop their own thinking. Parents and educators have welcomed the tool’s productivity benefits while worrying about over-dependence and hidden risks.

Those worries now have fresh data behind them. The institute’s report highlights how conversational AI differs from previous technologies. Its ability to maintain long, personalized dialogues creates unique hazards around emotional attachment and undetected distress. Traditional content filters cannot fully address a system designed to respond adaptively to whatever a user says.

Regulators and lawmakers have taken notice of similar concerns. Lawsuits against OpenAI have alleged that heavy engagement with the chatbot contributed to suicide attempts by young people. While causation remains disputed, the pattern of emotional reliance appears across multiple accounts.

For now, the Youth AI Safety Institute recommends that teens avoid ChatGPT entirely for emotional support or mental health topics. The tool may assist with certain learning tasks under close supervision. Even then, parents should not assume the advertised guardrails will protect them.

OpenAI faces pressure to respond with more than statements. Independent testing that confirms the fixes will be essential. Until then, the gap between marketing claims and real performance leaves families exposed. The chatbot’s popularity among teens only heightens the stakes.

Child development experts who reviewed the test responses expressed particular concern about extended conversations. A single missed alert or inadequate crisis reply can compound over minutes or hours. In mental health crises, those minutes matter.

The institute plans to continue its evaluations of AI products aimed at youth. Its standards remain independent even though it has received some philanthropic support from the OpenAI Foundation. That separation, officials stress, preserves the credibility of findings that companies may find uncomfortable.

Parents seeking alternatives have limited options. Many schools still grapple with how to incorporate or restrict generative AI. Some districts ban it. Others attempt to teach responsible use. Few have settled on clear policies that account for the latest safety data.

The Common Sense Media assessment arrives as a warning rather than a final judgment. Technology can improve. OpenAI has already iterated on safety features multiple times. The question is whether future changes will meet the rigorous bar set by child safety researchers — and whether those changes will be proven before millions more adolescents incorporate the tool into their daily lives.

So far, the evidence suggests caution. Families cannot afford to treat experimental guardrails as reliable protection. The risks, this latest analysis shows, remain too high.