AI Safety Claims Need An Expiration Date – OpEd
The Paris Call for AI Safety and Security commits signatories to shared frontier-model evaluations and a public repository, but the piece stresses that safety claims must travel with their assumptions, scope and invalidation conditions.
Formally verified results are mathematically rigorous only under stated boundaries; presenting a narrow proof as a general safety guarantee can mislead more than an ordinary benchmark.
A compact claim-boundary record—detailing the exact property, artifact, assumptions, independent scrutiny and what changes would invalidate the result—is proposed to make international reporting precise and trustworthy.
The new Paris agreement promises common evaluations. Global trust will depend on whether every safety claim also discloses its assumptions, failure boundaries and the changes that would invalidate it.
On November 9, governments and leading AI companies endorsed the Paris Call for AI Safety and Security. Its signatories committed to work toward shared frontier-model evaluations, more comparable reporting and a public repository of non-sensitive findings. They also emphasized independent assessment and disclosure of model capabilities, limitations and known risks. The declaration is a political commitment, not a global certification law. Still, it puts a neglected question directly on the international agenda: when a safety finding travels between countries, will its limitations travel with it?
The answer matters most when the finding sounds conclusive. “Formally verified” is one of the strongest phrases in engineering. In its proper setting, it describes a rigorous result: a specified property has been established for a mathematical model or implementation under stated assumptions. It does not mean that everything a connected AI system might do, in every operating environment, has been proved safe.
A convincing proof with an omitted boundary can be more misleading than an ordinary benchmark. People know to question a score. They may treat the word proved as the end of the argument.
The part of a proof that rarely makes the headline
Imagine a service that automatically acts on behalf of a human operator. Engineers formally check a model of its authorization mechanism and establish that, once permission is withdrawn, the model permits no further prohibited action. That can be a valuable theorem. But suppose the model represents withdrawal as instantaneous while the deployed service has queued commands, network delays and cached credentials. A real action may still complete after the withdrawal request. The mathematical result has........
