AI Act: ‘Every action is logged’, ‘a human validates every email’… under what conditions can these promises regarding agents be verified?

Section 12 of the AI Act requires events to be recorded without specifying which ones prove what. Researchers identify four factors that must be established before any assessment of a claim regarding an AI agent can be carried out.

Organizations are making more and more claims about their AI agents: one person approves every message sent outside, all actions are tracked, an evaluation has shown that the agent can be deployed safely. Article 12 of the EU AI Act requires high-risk systems to allow automatic recording of events, but does not specify which recordings are sufficient to confirm or refute a given claim.

Name the rule, scope, evidence, and author

Yue Zhao and his co-authors argue that in order to be verifiable, a claim must first designate the rule it concerns, its scope, the records that would make it possible to make a decision, and who produces them. Drawing on the prerequisites of an assurance engagement, they describe a claim as "auditable" when these elements and a decision criterion are set before any verdict is reached, and the records are accessible. Agents add three requirements: a record that is independent of the agent, an authorization tied to the specific arguments for each action, and a completeness that goes beyond the mere integrity of the logs.

The article applies the method to six common claims based on draft standards from NIST, IETF, and OWASP, tracks a case through five levels of evidence, and concludes with recommendations for operators, buyers, auditors, and standards organizations.