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·Aug 27·5 min read
Collective alignment: public input on our Model Spec
TL;DR
AI systems often reflect the values of their creators rather than diverse global perspectives, creating misalignment with broader human preferences. OpenAI conducted a survey of over 1,000 people worldwide to understand how AI should behave, then compared responses against their Model Spec—a document defining AI behavior defaults.
✦ Why It Matters
Engineers can use collective alignment methodology to validate AI behavior specifications against real user values before deployment.
Key Takeaways
- Survey 1,000+ users globally to identify gaps between AI design and public values before deployment.
- Compare user feedback directly against your Model Spec to find specification-level misalignments.
- Use demographic breakdowns to understand which groups have divergent preferences on safety and helpfulness.
- Iterate AI defaults based on collective input rather than designer assumptions alone.
Full Summary
OpenAI identified a gap between how AI systems are currently designed and what diverse populations globally believe AI should do. The company surveyed over 1,000 people across different regions to gather input on AI behavior preferences, ethics, and values.
This approach, called collective alignment, treats public input as a systematic input to AI design rather than an afterthought. Responses were compared directly against OpenAI's Model Spec—the technical specification document that defines default behaviors, safety constraints, and response patterns for their AI models.
The survey captured measurable differences in how various demographic groups prioritize safety, helpfulness, honesty, and other behavioral dimensions. Results showed specific areas where public consensus diverged from existing specifications, enabling engineers to adjust model training and deployment accordingly.
This methodology demonstrates how large-scale human feedback can inform AI defaults at the specification level, not just at deployment.