OpenAI Details Safety Alignment Stack, Reporting 97% Refusal on Uncertain Requests

OpenAI’s February 13, 2026 write-up on safety alignment provides a concrete operational framing: model behavior should be governed by structured instruction priority, not by flat prompt compliance. The document argues that many failures are not about missing knowledge, but about failing to resolve conflicting directives safely.

The proposed hierarchy orders instruction sources as system, developer, user, then guideline. Under this “chain of command,” lower-priority requests should only be followed when they do not conflict with higher-priority policies. This is meant to improve robustness against prompt injection and ambiguous conversations where intent is mixed.

OpenAI places special emphasis on uncertainty handling. In the company’s reported evaluation, a baseline strategy refused around 59% of uncertain requests while still answering around 41%. With uncertainty-aware behavior plus explicit command hierarchy, refusal rose to around 97%, abstentions were around 3%, and direct answers in uncertain cases dropped to effectively zero. The intended effect is to reduce unsafe over-compliance when policy interpretation is unclear.

The release also cites ConflictQA-style evaluations, including 95.4% rule-following and 100% correct abstentions in specific no-valid-answer conditions. OpenAI further reports large gains in synthetic command-conflict scenarios when hierarchy-aware reasoning is enabled. Together, these metrics suggest that “knowing when not to answer” is being treated as a first-class capability, not an afterthought.

For enterprise deployment, the implication is straightforward: safety should be measured with dual objectives. Teams need capability metrics for task completion and separate control metrics for abstention quality, refusal precision, and policy-consistent behavior under adversarial prompts.

Important caveats remain. Vendor benchmarks may not fully represent real multilingual attack surfaces, and stricter refusal policies can increase false refusals that degrade user experience. Organizations should therefore replicate these tests on internal workloads, tune policy layers by domain risk, and monitor refusal-error tradeoffs continuously. Even with those constraints, the paper is notable for translating safety alignment from principles into testable, operations-oriented design.