OpenAI released an in-depth public explanation of how its Model Spec works, how it is updated, and why it is shared openly. The document accompanies recent Spec changes, including a renamed top authority level — Root, formerly Platform — and expanded guidance on agentic safety as models move beyond chat.

Why it matters

Most AI companies treat behavioral guardrails as proprietary. OpenAI's decision to publish the Model Spec under CC0 — dedicated to the public domain — creates an external accountability layer that researchers, journalists, and regulators can actually use. When a model behaves unexpectedly, the Spec gives observers a reference point: was this consistent with OpenAI's stated intent, or a deviation? That distinction matters enormously as regulators in the EU and US begin auditing AI systems against published safety claims.

What's new in the Spec

The 2026 updates refine the authority hierarchy — now Root → System → Developer → User → Guideline — to better handle agentic settings where an AI might receive conflicting instructions from multiple principals. The Spec also expands guidance on model personality, safety-versus-helpfulness trade-offs, and how models should handle situations where user requests conflict with developer system prompts. A new section addresses teen protections, synchronized with OpenAI's gpt-oss-safeguard release.

Competitive context

Anthropic's Constitutional AI papers and model cards serve a similar transparency function, though they are framed more as research publications than operating policy. Google has published responsible AI practices but does not maintain a single behavioral specification in public-domain form. The Model Spec's CC0 license is unusual — it invites other labs to fork and adapt it, which could become the foundation for an industry-wide behavioral standard if regulators push for harmonization.