Back to OpenAIAll frontier AI labs
Ethics and safety policies
OpenAI: what they say, and what they leave out
Every claim on this page comes from documents OpenAI published itself, linked below so you can check me. Checked September 19, 2026.
A policy is not an audit
Almost every document linked on this page was written by the company it governs, and graded by that same company. A long policy list does not make a lab more ethical than a short one. This page keeps three things apart: what the lab has committed to in writing, what its documents do not cover or where they have been credibly criticized, and — clearly labeled at the end — my own read.
Their stated position
OpenAI frames its mission as ensuring artificial general intelligence benefits all of humanity, and publishes both a rulebook for how its models should behave and a framework for deciding when a capability is too dangerous to release.
It positions safety as something that improves through staged deployment: release to a limited group, watch what happens, widen access.
The documents themselves
Read these rather than this summary if a decision depends on it.
- Model Spec
The intended behavior of its models, including the 'red line' principles it says should never be crossed.
- Preparedness Framework
How it categorizes dangerous capabilities — cyber, bio, self-improvement — and what thresholds trigger extra safeguards.
- Usage policies
What users are forbidden from doing with the products.
- Safety and responsibility hub
Its public summary of safety work, system cards and evaluations.
What they have committed to
Stated in their own published policies, in writing.
- — Publishes a system card for major model releases, describing evaluations run before launch.
- — Commits to withholding or restricting models that cross its own stated capability thresholds.
- — States in the Model Spec that some behaviors are off-limits regardless of user instruction.
What those documents don't cover
Plain absences, and criticism that has been documented publicly — not speculation.
- — Every threshold in the Preparedness Framework is set, measured and graded by OpenAI itself; no external body can compel a delay.
- — The framework has been revised more than once, and critics have noted revisions loosened language on some categories.
- — Training-data sources are not disclosed. Multiple copyright suits are unresolved.
- — Several senior safety researchers left publicly in 2024 saying safety work lost internal priority — the company disputes that characterization.
My read
This section is my opinion, clearly separated from everything above. Disagree with it freely.
The Model Spec is useful reading and more specific than most labs' equivalents. It is still a description of intent written by the party it constrains, and intent is not an audit.