In its latest effort to address growing concerns about the impact of AI on young people, OpenAI on Thursday updated its guidelines on how its AI models should operate with users under 18 and published new AI literacy resources for teens and their parents. Still, questions remain about how consistently these policies are translated into practice.
The update comes as the AI industry in general, and OpenAI in particular, faces increased scrutiny from policymakers, educators, and child safety advocates following the alleged suicide deaths of several teenagers after lengthy conversations with AI chatbots.
Gen Z, which includes those born between 1997 and 2012, are the most active users of OpenAI’s chatbots. And after OpenAI’s recent deal with Disney, even more young people may flock to a platform that can do everything from asking for help with homework to generating images and videos on thousands of topics.
Last week, 42 state attorneys general signed a letter to Big Tech companies urging them to implement safeguards in AI chatbots to protect children and vulnerable populations. And as the Trump administration considers what federal standards for regulating AI will look like, policymakers like Sen. Josh Hawley (R-Missouri) have introduced legislation that would outright ban minors from interacting with AI chatbots.
OpenAI’s updated model specification provides operational guidelines for large-scale language models and builds on existing specifications that prohibit models from producing sexual content involving minors or encouraging self-harm, paranoia, or mania. This works in conjunction with an upcoming age prediction model that identifies when an account belongs to a minor and automatically deploys safeguards for teens.
Stricter rules apply when teenagers use models compared to adult users. Models are instructed to avoid immersive romantic role-play, first-person intimacy, and first-person sexual or violent role-play, even if it is not graphic. The specification also calls for special attention on topics such as body image and disordered eating, instructs models to prioritize communication about safety over autonomy when harm is involved, and to avoid advice that helps teens hide risky behavior from their parents.
OpenAI specifies that these limits should be maintained even when prompts are framed as “fictional, hypothetical, historical, or educational,” i.e., common tactics that rely on role-playing or edge-case scenarios to force AI models to deviate from guidelines.
tech crunch event
san francisco
|
October 13-15, 2026
Actions speak louder than words.
OpenAI says its key safety measures for teens are underpinned by four principles that guide the model’s approach.
We prioritize the safety of youth, even when other users’ interests, such as “maximum intellectual freedom,” conflict with safety concerns. Facilitate real-world support by directing teens to family, friends, and local professionals to bring them well-being. Rather than belittling teens or treating them like adults, talk to them with warmth and respect and treat them like teens. Be transparent by explaining what the assistant can and cannot do, and remind teens that the assistant is not a person.
The document also provides several examples of chatbots that explain why chatbots cannot “role-play as a girlfriend” or “assist with extreme appearance changes or dangerous shortcuts.”
Lily Lee, a privacy and AI lawyer and founder of Metaverse Law, said it’s heartening to see OpenAI taking steps to ensure its chatbots don’t engage in such behavior.
Explaining that one of the biggest complaints that advocates and parents have about chatbots is that they relentlessly encourage continued engagement in a way that can be addictive for teens, she said, “I’m really happy to see that OpenAI is saying in some of these answers that it can’t answer your question. I think the more we know about it, the more it will break the cycle that leads to a lot of inappropriate behavior and self-harm.”
That said, the examples are just that: a cherry-picked version of how OpenAI’s safety team wants the model to behave. Pandering, the tendency for AI chatbots to over-agree with users, was listed as a prohibited behavior in an earlier version of the model specification, but ChatGPT still engaged in that behavior. This is especially true for GPT-4o, a model that has been linked to some instances of what experts call “AI psychosis.”
Robbie Torney, senior director of AI programs at Common Sense Media, a nonprofit organization dedicated to protecting children in the digital world, expressed concern about potential inconsistencies within Model Spec’s under-18 guidelines. He highlighted the tension between safety-focused regulations and the “no topic is off limits” principle, which directs models to address any topic, regardless of sensitivity.
“We need to understand how the different parts of the specification fit together,” he said, noting that certain sections could move the system toward engagement over safety. He said his organization’s testing has found that ChatGPT often reflects the energy of its users, sometimes resulting in responses that are not situationally appropriate or in line with user safety.
Their conversations illustrate the case of Adam Lane, a teenage boy who died by suicide after months of interaction with ChatGPT, a chatbot that was conducting such mirroring. The incident also revealed that OpenAI’s moderation API failed to prevent unsafe and harmful interactions, despite flagging more than 1,000 instances of ChatGPT that referenced suicide and 377 messages that included self-harm content. But that wasn’t enough to stop Adam from continuing the conversation with ChatGPT.
In an interview with TechCrunch in September, former OpenAI safety researcher Steven Adler said this is because OpenAI historically ran classifiers (automated systems that label and flag content) in batches after the fact rather than in real time, and did not properly gate user interactions with ChatGPT.
OpenAI now uses automated classifiers to evaluate text, image, and audio content in real-time, according to the company’s updated parental control documentation. The system is designed to detect and block content related to child sexual abuse, filter sensitive topics, and identify self-harm. If the system flags a prompt that indicates a serious safety concern, a small trained team may review the flag to determine if there are signs of “acute distress” and notify parents.
Torney praised OpenAI’s recent efforts on safety, including its transparency in publishing guidelines for users under 18.
“Not all companies publish their policy guidelines in the same way,” Toney said, pointing to the leak of Meta’s guidelines that showed Meta had its chatbots having sensual and romantic conversations with children. “This is an example of transparency that helps safety researchers and the public understand how these models actually work and how they are supposed to work.”
But what ultimately matters is how the AI system actually works, Adler told TechCrunch on Thursday.
“We appreciate that OpenAI is thoughtful about intended behavior, but unless companies measure actual behavior, intent is ultimately just words,” he said.
In other words, what is missing from this announcement is any evidence that ChatGPT actually follows the guidelines set out in the model specification.
paradigm shift
Experts say these guidelines appear to set OpenAI up to pre-empt certain laws, such as California’s recently signed bill SB 243 regulating AI companion chatbots, which goes into effect in 2027.
The new language in the model specification reflects some of the law’s key requirements for chatbots to prohibit conversations about suicidal ideation, self-harm, or sexually explicit content. The bill would also require platforms to provide minors with an alert every three hours reminding them that they are talking to a chatbot rather than a real person and that they should take a break.
When asked how often ChatGPT will remind teens that they’re talking to a chatbot and remind them to take a break, an OpenAI spokesperson didn’t provide details, only saying that the company trains its models to represent and remind users as AI, and that it implements break reminders during “long sessions.”
The company also shared two new AI literacy resources for parents and families. These tips include conversation starters and guidance to help parents talk to their teens about what AI can and cannot do, building critical thinking, setting healthy boundaries, and navigating sensitive topics.
Taken together, these documents formalize an approach that shares management and responsibility. OpenAI provides families with a framework that details what the model is supposed to do and oversees how it is used.
The focus on parental responsibility is notable because it reflects Silicon Valley talking points. In its recommendations for federal AI regulation published this week, venture capital firm Andreessen Horowitz proposed more disclosure requirements for child safety, rather than restrictive requirements, and an emphasis on parental responsibility.
Some of OpenAI’s principles – When values conflict, safety first. Direct users to real-world support. It emphasizes that chatbots are not humans and is presented as a guardrail for teenagers. However, several adults have died by suicide and suffered from life-threatening delusions, requiring due follow-up. Should these defaults be applied across the board, or does OpenAI view them as a trade-off that it only wants to enforce when minors are involved?
An OpenAI spokesperson countered that the company’s safety approach is designed to protect all users, and said model specs are just one part of a multi-layered strategy.
Lee said it’s been “a bit of a Wild West” so far in terms of legal requirements and the intentions of technology companies. But she feels laws like SB 243, which would require tech companies to publish their safety measures, would change the paradigm.
“If a company advertises on its website that it has these safeguards and mechanisms in place, but fails to comply with putting these safeguards in place, there will be legal risks for the company,” Lee said. “Because from a plaintiff’s perspective, you’re not just looking at standard litigation and legal complaints, you’re also looking at potential unfair and deceptive advertising complaints.”
