Loading page…
A reasoning-based classifier that evaluates content against a supplied policy, making policy quality part of the evaluation.
This approach is relevant when a fixed moderation taxonomy does not express the rules of a product. Rather than assuming the model knows the intended boundary, the operator supplies a policy along with the content. The new flexibility makes ambiguity in that policy consequential.
OpenAI describes gpt-oss-safeguard as an open-weight model that takes a policy and content, then produces a classification with reasoning. Developers determine how to use that result in a pipeline.
Source: OpenAI · Introducing gpt-oss-safeguardThe published approach supports policies supplied at inference time. Our recommendation is to version policy text alongside model settings and evaluate revisions before rollout.
Source: OpenAI · Introducing gpt-oss-safeguardAn explanation gives a reviewer something to inspect. It does not establish that the model interpreted an exception correctly. Include examples that satisfy one clause while conflicting with another, and compare against human judgments.
A product-review forum wants to identify undisclosed promotional posts while allowing ordinary recommendations. Write examples for both, include ambiguous disclosures, and test whether policy edits fix those cases without blocking legitimate reviewers.
This is not a guarantee that any written policy will be followed, a factual verification service, or a measured latency comparison with smaller classifiers.
How to read an AI safety evaluationThe sources behind this page, with a reason to open each one. Practical examples and recommendations are our editorial interpretation.
Describes classification using a supplied policy and reasoning. Reported results are the provider’s evaluations.
Pairs safe prompts with unsafe contrasts to investigate unnecessary refusals. Historical model results are not current rankings.
Sources reviewed 13 September 2026. Product documentation can change. How we use evidence