EU AI Act: Article 14 human oversight
Evidence that natural persons effectively oversee high-risk AI systems as required by Regulation (EU) 2024/1689. Each attestation is a per-output record that an identified human understood, reviewed, and where necessary intervened in the system's output.
Pack metadata
The default clause applies to review requests that name no clause and match no scope. That is what makes activating the pack organization-wide plug-and-play.
Clauses
Each clause maps a legal obligation to the attestation evidence that satisfies it: an allowlist of accepted review kinds, a minimum signer tier, and whether the attestation must carry a named signer role.
| Reference | Clause | Obligation | Accepted review kinds | Min signer tier | Named role |
|---|---|---|---|---|---|
| EU AI Act, Art 14(1), 14(4)(d)-(e) | Effective human oversight of high-risk AI outputart14-oversight | High-risk AI systems must be designed so natural persons can effectively oversee them, including the ability to decide not to use an output, to disregard or reverse it, and to intervene or interrupt operation. A rejection attestation is itself evidence of an exercised intervention. | reviewed, revised, endorsed, rejected | SSO verified | Required |
| EU AI Act, Art 14(4)(c) | Correct interpretation before relianceart14-interpret | Persons overseeing the system must be able to correctly interpret its output before it is relied on, taking account of the system's capabilities, limitations, and automation bias.scopes: high-risk-decision, interpretation | reviewed, revised | Org attested | Required |
Immutability
This published version never changes. The canonical JSON is what attestations reference via policy_uri, and it resolves, byte-identical, forever. Amendments ship as a new version of the pack, published alongside this one; see governance.