We’re launching Lyria 3.5 in Google Flow Music, with advances across musicality, lyrics, vocals, and creative control
deepmind.google·6d ago
AI Engineering
OpenAI Blog·Dec 18·5 min read
Updating our Model Spec with teen protections
TL;DR
ChatGPT lacked age-appropriate safety guidelines tailored to teen developmental needs, creating gaps in how the model handles sensitive topics with younger users. OpenAI updated its Model Spec—the document defining model behavior—with new Under-18 Principles grounded in developmental science to provide safer, age-appropriate guidance.
✦ Why It Matters
Engineers can now implement age-aware safety constraints in conversational AI systems using developmental science principles.
Key Takeaways
- Incorporate developmental science research when designing safety rules for teen-facing AI systems.
- Update model specifications to explicitly define behavior for different age groups, not just adults.
- Strengthen guardrails in high-risk scenarios like mental health or peer pressure situations.
- Test age-appropriate responses separately from general model behavior to catch developmental mismatches.
Full Summary
OpenAI identified that ChatGPT's existing safety guidelines did not adequately account for the developmental differences between teens and adults, leaving gaps in age-appropriate support. The company updated its Model Spec—a technical specification document that defines how AI models should behave across different scenarios—by adding Under-18 Principles.
These principles are grounded in developmental science, meaning they incorporate research on how adolescent brains develop cognitively and emotionally. The update strengthens guardrails (safety constraints that prevent harmful outputs) and clarifies expected model behavior in higher-risk situations such as mental health crises, peer pressure, or sensitive personal topics.
This change represents part of OpenAI's broader initiative to improve teen safety across ChatGPT. The concrete outcome is a more nuanced set of behavioral rules that allow the model to provide supportive, developmentally appropriate responses rather than generic adult-oriented guidance.