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·Sep 2·5 min read
Building more helpful ChatGPT experiences for everyone
TL;DR
ChatGPT lacked adequate safeguards for sensitive conversations and teen users, creating risks around harmful content and age-appropriate interactions. OpenAI implemented three concrete changes: expert partnerships for safety review, parental controls for minors, and routing complex ethical queries to reasoning models that think through responses step-by-step.
✦ Why It Matters
Engineers can adopt expert-in-the-loop review and reasoning model routing to improve safety in production AI systems.
Key Takeaways
- Integrate domain experts into safety review cycles before deploying consumer-facing AI features.
- Implement routing logic that directs high-stakes queries to reasoning models with explicit step-by-step outputs.
- Add parental oversight mechanisms for products serving minors to enable guardian control.
- Measure safety improvements through user satisfaction and harmful content reduction metrics.
Full Summary
ChatGPT faced growing pressure to handle sensitive topics—conversations involving mental health crises, abuse, or age-inappropriate content—without sufficient guardrails. Teens using the platform had no parental oversight mechanisms, and the base model sometimes generated harmful or inconsistent responses to ethically complex questions.
OpenAI addressed this through three coordinated initiatives: partnering with domain experts (psychologists, safety researchers, child development specialists) to audit and improve response quality; deploying parental controls that let guardians monitor teen accounts and set usage boundaries; and implementing a routing system that directs nuanced ethical queries to reasoning models—advanced AI systems that explicitly work through multiple perspectives before responding. The reasoning models use chain-of-thought processing, showing their reasoning steps rather than jumping to conclusions.
Results included reduced harmful content in sensitive domains and improved user satisfaction scores among teen and guardian populations. These changes establish precedent for responsible AI deployment in consumer applications.