EU AI ActAI Act human oversight evidence

Human oversight evidence

Record approvals, escalations and overrides so human oversight remains provable after deployment.

Last reviewed: 2026-06-03Official source links includedRobots: index,followQuality gate: approved
Relevant evidence visual for Human oversight evidence

What this page covers

Human oversight is credible only when the reviewer, system state, decision and timestamp remain connected.

Legal timing

The AI Act entered into force on 1 August 2024. The Commission describes 2 August 2026 as the general application date, while the AI Omnibus political agreement moves many high-risk areas to 2 December 2027 and product-integrated systems to 2 August 2028. The preparation work still has to start now because evidence cannot be rebuilt reliably after deployment.

This page is implementation guidance for evidence planning, not legal advice.

Evidence Attesto AI can preserve

human approval or override

system identifier

model or service version

policy version

auditor verification receipt

Example evidence records

human approval or overridetrusted timestamprisk control resultcanonical event hashMerkle inclusion proof

Example proof receipt

Example Attesto receipt

event_type

HUMANOVERSIGHT

timestamp

2026-06-04T10:21:00Z

leaf_hash

sha256:8f41...b19e

merkle_root

sha256:52ac...91d4

verification_status

valid demo receipt, raw data not exposed

Where Attesto fits

Attesto ties human decisions to hashes, policy versions and verifier receipts instead of leaving them buried in chat or tickets.

FAQ

How is this different from a normal log?

A normal log asks an auditor to trust the system that produced it. Attesto records hashes, signatures, Merkle proofs and verifier receipts so selected evidence can be checked independently.

Does Attesto need to expose raw sensitive data?

No. Raw records can remain encrypted or customer-controlled while proof material is shared for verification.

Where does Human oversight evidence fit in the compliance stack?

Attesto ties human decisions to hashes, policy versions and verifier receipts instead of leaving them buried in chat or tickets.