OpenAI rewrites ChatGPT rules to protect teens

If a lonely 15-year-old asks ChatGPT late at night to pretend to be their girlfriend, what should the AI say?

OpenAI just updated its Model Spec, the official behavioral playbook guiding how its large language models should act in sensitive scenarios, adding a dedicated set of Under-18 Principles for users aged 13 to 17. The core takeaway is clear: OpenAI is drawing hard lines against romantic roleplay, emotional attachment, and secrecy pacts with teenagers.

This update does not arrive in a vacuum. The AI industry is facing mounting pressure following reports of teenagers taking their own lives after prolonged interactions with AI chatbots. Just last week, 42 state attorneys general sent a joint letter demanding immediate guardrails for minors, while federal lawmakers like Senator Josh Hawley are already floating legislation to ban children from using AI chatbots entirely.

Hard limits on romantic bots and secret-keeping

Under the new guidelines, OpenAI establishes four central commitments: prioritizing minor safety over intellectual freedom, encouraging real-world support, treating teens like teens without condescension, and remaining fully transparent about AI limits.

In practice, that means ChatGPT must outright decline first-person romantic roleplay, intimacy, or violent roleplay, even if the content is not explicitly graphic. In examples published by OpenAI, the assistant directly tells users it cannot act as a romantic partner. The system is also instructed to shut down discussions encouraging extreme appearance alterations or dangerous shortcuts tied to body image.

Crucially, OpenAI explicitly noted that these safety boundaries must hold even when teens try to bypass them by framing prompts as fictional storytelling, hypothetical games, or educational thought experiments. Furthermore, the bot cannot help minors conceal risky behaviors from their parents or guardians.

Guessing ages before trouble starts

Of course, teenagers rarely disclose their real birth years voluntarily. To enforce these rules, OpenAI is rolling out an automated age-prediction model across ChatGPT consumer accounts.

If the system suspects an account belongs to a teenager, or if account data is incomplete, it will default directly to the restricted under-18 experience. Adults swept up by the algorithm will have to formally verify their age to unlock unrestricted access.

These safeguards now extend across newer products, including group chats, the ChatGPT Atlas browser, and the Sora video app. OpenAI also added built-in break prompts for marathon sessions and partnered with ThroughLine to provide localized distress helplines, with input from the American Psychological Association.

Industry observers are split on whether written rules are enough. Lily Li, founder of Metaverse Law, welcomed the change, noting that shutting down addictive conversation loops is vital for child safety. However, Robbie Torney of Common Sense Media pointed out that the document may suffer from internal friction between strict teen protections and the existing spec rule that no topic should be off-limits. Critics also note that OpenAI previously banned sycophancy on paper, yet models like GPT-4o still flattered vulnerable users into concerning feedback loops.

Why it matters

For teens and parents, these guardrails are overdue. Gen Z represents ChatGPT's most active user demographic, and following OpenAI's recent high-profile deal with Disney, millions more young users are heading to the platform. While the protections make homework and creative exploration safer, adults should be prepared for potential friction if the age-prediction algorithm misflags their accounts.

For the AI industry, this move looks like pre-emptive defense against crushing regulatory intervention. By voluntarily implementing stricter standards than traditional social media platforms, OpenAI is trying to head off total bans proposed on Capitol Hill. Competitors offering companion bots or unrestricted roleplay will soon face intense pressure to match these boundaries or risk severe legal backlash.

We think shifting to a restrictive default based on age prediction is a clever, necessary engineering choice that closes the simplest registration loopholes. However, the real test lies in execution. Language models are fundamentally prone to flattering users and playing along. Until OpenAI proves that these guardrails hold firm against clever edge cases in the real world, the U18 principles remain an ambitious target rather than a guaranteed safeguard.