Abstract

This disclosure describes a registry that makes the certification of AI systems itself checkable. The registry lists every body accredited to certify AI systems and every laboratory accredited to test them, with the scope of each accreditation, its live status, and the keys each body uses to sign. Every certificate, conformance level and passport status issued by a body is logged by the registry in a public, append-only log, and every status carries a reference to the issuing body's accreditation, so that a verifier checking an AI system's passport also checks, in the same step, that the body vouching for it is in good standing. The registry tests the testers: it submits seeded systems, built with known, hidden flaws, to certification bodies and laboratories as if they were ordinary applicants, and compares their findings with the flaws seeded; it also re-tests a random sample of systems each body has certified. A body that misses seeded flaws, or whose certified systems fail re-testing, has its accreditation narrowed, lowered or suspended automatically under published rules, and its status changes at once in every chain that relies on it. When a body loses accreditation, the systems it certified are not cut off at once: they enter a declared transfer period during which they keep a marked status and must be re-certified by another accredited body, chosen without the original body's involvement. Rules on independence, such as limits on fees from any one maker and rotation of assessors, are recorded and checked. Registries in different regions that apply the same rules recognize one another, and the registry is mirrored by independent parties.

Creative Commons License

Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.

Download

Share

COinS