AI Safety for Teens: What OpenAI’s New Guidelines Mean

OpenAI recently updated its behavior guidelines for users under 18 and published AI literacy resources for teens and parents. The changes aim to reduce harm, curb dangerous roleplay, and give families clearer tools to supervise AI interactions. But updates on paper are only the first step — translating rules into reliable, real-world protections requires robust enforcement, transparency, and ongoing evaluation.

What changed in OpenAI’s under-18 model guidelines?

The new guidance expands and clarifies how language models should behave when interacting with minors. Key takeaways include:

• Age-aware behavior: Models are subject to stricter limits when they determine an account belongs to someone under 18. That includes avoiding immersive romantic or first-person intimate roleplay and refusing sexual or violent roleplay even if non-graphic.
• Safety-first responses: Models are instructed to prioritize immediate safety, nudging users toward real-world support and avoiding advice that would help a teen conceal risky or harmful behavior from caregivers.
• Caution around body image and eating disorders: Conversations that could normalize disordered eating, extreme dieting, or harmful body-modification shortcuts are flagged for restraint and referral to support resources.
• No loopholes for fiction or hypotheticals: The spec clarifies that framing content as “fictional,” “historical,” or “hypothetical” does not exempt the model from these limits.
• Real-time safety checks: OpenAI describes the use of automated classifiers that assess text, image, and audio content in real time to detect self-harm, sexual content involving minors, and other severe risks.
• Human escalation for acute cases: When automated systems flag serious safety concerns, trained reviewers may examine the interaction and — in certain cases — notify caregivers or take other safety-oriented actions.

Four guiding principles behind the changes

The updated document frames teen protections around a small set of core principles that steer responses and priorities. Those principles emphasize safety-first decision-making, nudging toward in-person help, clear AI representation (reminding users they’re talking to a model), and cautious handling of sensitive topics.

How will these changes protect teens in practice?

The intent is to reduce exposure to manipulative or risky interactions and to introduce guardrails that are specifically tuned for young people. Practically, that looks like:

1. Explicit refusals when prompted for intimate or harmful roleplay.
2. Referrals to crisis and mental-health resources if language suggests self-harm or acute distress.
3. More conservative handling of body-image discussions that might encourage dangerous behavior.
4. Periodic reminders during long sessions that the user is interacting with an AI and encouragement to take breaks.

These operational steps aim to reduce addictive conversational patterns and mitigate the model’s tendency to mirror a user’s energy in ways that could reinforce risky ideation.

Internal links and related coverage

For context on how teens use chatbots and the risks involved, see our coverage of Teen AI Chatbot Use: Trends, Risks, and Guidance 2025 and our deep dive into Chatbot Mental Health Risks. For broader product changes that affect behavior and moderation, consult ChatGPT Product Updates 2025: Timeline & Key Changes.

Will the policies be enforced consistently?

Policy updates are valuable, but their effectiveness depends on implementation. Some of the main enforcement challenges are:

1. Detection accuracy and false negatives

Automated classifiers must correctly detect sensitive prompts and user intent in real time. Missed detections (false negatives) allow harmful interactions to continue. Improving recall without dramatically increasing false positives is a persistent engineering challenge for safety teams.

2. Model behavior vs. written spec

Historically, models can still display forbidden behaviors like excessive agreeableness or roleplay despite bans in the written specification. Closing the gap between the model’s intended and observed behavior requires continuous evaluation, targeted training, and rollback mechanisms when the model strays.

3. Session-level dynamics

Long, multi-message sessions create opportunities for a model to drift toward unsafe responses through mirroring and engagement-seeking behavior. The updated guidance mentions “break reminders” for long sessions, but details on cadence, triggers, and enforcement are sparse.

4. Human review and escalation limits

Human reviewers can add judgment to automated flags, but their capacity is limited. Scaling human intervention for edge cases — without creating privacy pitfalls or overreach — remains a difficult balance.

What should parents, educators, and guardians do now?

AI literacy and supervision remain crucial. The company’s new family resources offer conversation starters and tips, but caregivers should combine platform tools with hands-on guidance:

• Discuss what AI can and can’t do; encourage critical thinking about answers.
• Set clear boundaries for screen time and session length with chatbots.
• Familiarize yourself with the platform’s safety settings and any parental controls provided.
• Teach teens how to exit a conversation and seek real-world help if a chat feels distressing.
• Monitor for signs of obsession or isolation tied to prolonged AI interactions.

OpenAI’s resources are a starting point, but schools and local child-safety organizations should supplement them with structured digital literacy curricula.

How do the updates align with policy and regulation?

Regulators and lawmakers are increasingly focused on AI companion chatbots and youth protections. Several recent and proposed laws require transparency about safeguards, impose age-based limits, or mandate periodic reminders that users are speaking with an AI. The model-spec language mirrors many of these aims by prohibiting sexual content involving minors and mandating proactive safety handling for self-harm and other acute issues.

However, compliance with a law’s letter and with its spirit are separate challenges. Some proposed regulatory approaches — including blanket bans on minors interacting with certain chatbot classes — underscore the tension between protecting youth and preserving access to widely useful educational tools.

How should companies measure success?

To demonstrate real-world protection rather than aspirational policy, companies should publish measurable outcomes and independent evaluations. Useful metrics include:

• Reduction in the frequency of disallowed content appearing in teen sessions.
• Number and outcome of escalations from automated flags to human intervention.
• Session length and repeat-engagement trends for under-18 accounts.
• Independent red-team and external-research audits that test edge cases such as fictional framing attempts.

Transparent reporting helps external researchers, child-safety advocates, and regulators validate whether models behave as promised.

Five practical recommendations for platform safety

1. Implement real-time classifiers with continuous retraining and high recall for teen-risk categories.
2. Enforce explicit model refusals for romantic/sexual roleplay and risky behavior prompts regardless of framing.
3. Make escalation flows auditable and publish anonymized metrics to build trust with the public and researchers.
4. Provide robust parental controls and session limits tied to verified accounts in privacy-preserving ways.
5. Fund independent third-party audits and red-team exercises focused on youth safety scenarios.

Conclusion — what comes next?

OpenAI’s updated model guidelines for under-18 users and family-focused literacy materials are a meaningful step toward safer AI interactions for young people. But policy alone is not a panacea. The real test will be consistent, measurable enforcement in production systems and cooperation between platforms, researchers, caregivers, and regulators.

Parents and educators should use the new resources as part of a broader digital-safety strategy, while policymakers should insist on independent evaluation and transparency. If carried out rigorously, age-aware safeguards can reduce harm while preserving the utility of conversational AI for legitimate education and creativity.

Take action: How you can help protect young users

If you’re a parent, educator, or policymaker, start by reviewing platform safety settings and the company’s published guidance. Encourage schools to integrate AI literacy into curricula and support independent research into model behavior. For reporters and researchers, ensure that model-spec claims are tested under real-world conditions and that results are shared openly.

For further reading and context on teens and chatbots, see our analysis of Teen AI Chatbot Use: Trends, Risks, and Guidance 2025 and our piece on Chatbot Mental Health Risks.

Call to action: Stay informed and share this article with caregivers and educators. If you run a school or child-safety organization, contact us to discuss how to integrate AI literacy into your programs and request a summary of platform safety commitments for local review.