Public explainer
Assurance Is Not a Safety Guarantee
Released July 17, 2026 · Current
Light-tier public explainer. Noncanonical. Authority effect: None. This piece is a readable derivative of Standards Body Terminology, Version 1.0, which governs if anything here differs from it. It is not an adopted standard and does not certify, accredit, or approve anything.
The word assurance sounds like a promise. In evidence-based practice it is something more modest and more useful: a structured engagement in which an independent practitioner evaluates evidence and provides a conclusion or defined findings. The value of assurance comes from its boundaries, and it is destroyed by pretending those boundaries do not exist.
Assurance comes in levels. Limited assurance provides moderate confidence based on narrower evidence and procedures. Reasonable assurance provides high but not absolute confidence, supported by more extensive evidence. Absolute assurance is, in Standards Body usage, prohibited as a practical claim, because no evaluation or audit can eliminate all uncertainty. Any claim of total assurance about a frontier AI system is a warning sign about the claimant, not information about the system.
Two distinctions keep assurance honest. First, capability is not risk: risk is the combination of the likelihood and consequence of an adverse outcome under a defined context, and what a system can do is not by itself what it will do, to whom, under which conditions. Second, passing is not proof of safety. An evaluation result is evidence about performance on what was actually measured, under the conditions actually used, at the time of measurement. It supports bounded inferences. It does not establish that a system is safe in general, and no result should be read that way.
This is why serious evaluation reporting states its scope, conditions, elicitation effort, limitations, and expiration, and why results should be treated as dated evidence rather than permanent properties. A system changes, its deployment context changes, and the world it operates in changes. Evidence that was worthy of reliance last year may not be this year.
None of this diminishes assurance. Bounded confidence, honestly stated, is exactly what decision-makers need and almost never get. The alternative to overclaiming is not silence. It is saying precisely what the evidence shows, for how long, and no more. That discipline is what separates assurance from reassurance.
Read next
- An Evaluation Is Not an Audit
- Static Benchmarks Are Not Enough, the public essay
- Standards Body Evidence Standards, the canonical source
Corrections: [email protected]. Corrections are recorded, and released versions are preserved without silent edits.