OpenAI has updated its safety guidelines for how ChatGPT should interact with users under 18, alongside publishing new AI literacy resources for teens and parents. The changes arrive as lawmakers, educators, and child-safety advocates intensify pressure on the AI industry over concerns that chatbots can negatively affect young people, particularly around mental health.

The move follows a series of high-profile cases in which teenagers allegedly died by suicide after extended interactions with AI chatbots, prompting renewed debate about whether existing safeguards are sufficient and whether regulation should go further.

Gen Z, broadly defined as people born between 1997 and 2012, represents the most active cohort of ChatGPT users. OpenAI’s expanding partnerships and use cases, from homework help to image and video generation, are likely to draw even more young users onto the platform.

At the policy level, pressure is mounting. Last week, attorneys general from 42 U.S. states urged major tech companies to implement stronger protections for children and vulnerable users. At the federal level, proposals are emerging that would sharply limit or even ban minors’ access to AI chatbots. Senator Josh Hawley (R-MO), for example, has introduced legislation that would prohibit minors from interacting with such systems entirely.

Stricter rules for teen users

The updated Model Spec from OpenAI builds on existing prohibitions against generating sexual content involving minors or encouraging self-harm, delusions, or manic behavior. It is designed to work alongside an upcoming age-prediction system that aims to identify underage users and automatically apply teen-specific safeguards.

Compared with adults, teen users are subject to tighter restrictions. The models are instructed to avoid immersive romantic roleplay, first-person intimacy, and any first-person sexual or violent roleplay, even if non-graphic. They must also apply extra caution when discussing topics such as body image and eating disorders, prioritize safety over autonomy when there is risk of harm, and avoid offering advice that could help teens hide unsafe behavior from parents or caregivers.

Notably, these limits apply even when prompts are framed as fictional, hypothetical, historical, or educational, a common tactic used to push models toward edge cases that bypass safeguards.

Four guiding principles

OpenAI says its teen safety approach is guided by four core principles:

Put teen safety first, even when it conflicts with goals like maximum intellectual freedom
Encourage real-world support by directing teens toward family, friends, or local professionals
Treat teens like teens, with warmth and respect rather than condescension or adult framing
Be transparent about the system’s limits and consistently remind users it is not human

The documentation includes sample responses showing the chatbot declining requests such as roleplaying as a romantic partner or assisting with extreme appearance changes or risky shortcuts.

Praise, but also skepticism

Lily Li, a privacy and AI lawyer and founder of Metaverse Law, welcomed the clearer boundaries, noting that one of the biggest concerns among parents and advocates is that chatbots can encourage compulsive engagement.

“I’m very happy to see OpenAI explicitly say, ‘We can’t answer your question,’” Li said. “The more we see that, the more it breaks the cycle that can lead to inappropriate conduct or self-harm.”

Still, experts caution that published examples reflect intentions, not necessarily real-world behavior. Past versions of the Model Spec prohibited sycophancy, the tendency for chatbots to be overly agreeable, yet users continued to experience it in practice. This behavior has been linked by some researchers to severe psychological effects, sometimes described as “AI psychosis.”

Robbie Torney, senior director of AI programs at Common Sense Media, also raised concerns about internal tensions within the guidelines. He pointed to potential conflicts between safety-focused rules and broader principles like “no topic is off limits,” which can push systems toward engagement rather than restraint. Testing by his organization found that ChatGPT often mirrors users’ tone and emotional intensity, occasionally producing responses that are misaligned with safety needs.

Real-time moderation and parental controls

OpenAI says it now relies on automated classifiers that assess text, image, and audio content in real time, rather than flagging issues after the fact. These systems are designed to detect child sexual abuse material, identify self-harm signals, and filter other sensitive topics. When content suggests serious risk, trained reviewers may assess whether there are signs of acute distress and, in some cases, notify a parent.

Torney praised OpenAI for publishing its under-18 guidelines publicly, contrasting this transparency with reports that other platforms have allowed romantic or sensual chatbot interactions with children.

Regulation and what comes next

Experts say the updated policies position OpenAI ahead of emerging laws such as California’s SB 243, which regulates AI companion chatbots and takes effect in 2027. That law prohibits chatbots from engaging in conversations about suicidal ideation, self-harm, or sexually explicit content with minors, and requires regular reminders that users are interacting with an AI, not a human.

OpenAI has not disclosed how frequently ChatGPT will issue such reminders, saying only that models are trained to represent themselves as AI and that break reminders appear during long sessions.

The company has also released new AI literacy resources for parents, offering conversation starters and guidance on setting boundaries, encouraging critical thinking, and navigating sensitive topics with teens.

Taken together, these steps formalize a shared-responsibility approach: OpenAI defines how the models should behave, while families are encouraged to actively supervise use. Critics note that this emphasis mirrors Silicon Valley’s broader stance, which often favors disclosure and parental responsibility over strict access limits.

A lingering question remains. Several of the safeguards now framed as teen-specific, such as prioritizing safety over engagement and reinforcing that the chatbot is not a person, could arguably benefit adults as well. Whether OpenAI will extend these defaults more broadly is still unclear.

As former OpenAI safety researcher Steven Adler put it, intentions matter less than outcomes. Without evidence that ChatGPT consistently follows the rules laid out in the Model Spec, the effectiveness of these new protections remains an open question.

Source