Article 50 creates several different evidence problems.

Article 50 is not one generic disclosure obligation. Depending on the organization’s role and the system involved, the review may concern direct AI interaction notices, machine-readable marking of synthetic content, disclosure of emotion recognition or biometric categorisation, deepfake labelling, or publication of AI-generated text on matters of public interest.

The first review question is therefore not “Do we have an AI label?” It is “Which Article 50 obligation applies to this provider, deployer, system, content type, and deployment context?”

Assessment boundary: AGICOMPLY does not determine legal applicability or certify Article 50 compliance. The ATO Readiness Baseline organizes supplied evidence for a defined review context and records where applicability, implementation, or support remains unresolved.

1. Evidence of role, scope, and applicability

A reviewer needs enough documentation to understand why a particular transparency measure was selected. Representative artifacts include:

  • the AI system inventory entry and current system version;
  • provider or deployer role analysis;
  • the affected interaction or content type;
  • deployment geography and intended audience;
  • documented exceptions or transitional treatment;
  • the responsible product, legal, governance, and technical owners.

Without this scope record, a notice or marking specification may be technically real but disconnected from the obligation it is intended to support.

2. Evidence of interaction disclosure

For systems that interact directly with people, evidence should show more than the approved wording. The package should preserve where the notice appears, when it is presented, which system version uses it, and who approved the implementation.

Representative artifacts

  • approved disclosure language;
  • interface specification and placement record;
  • first-interaction timing evidence;
  • accessibility and localisation review;
  • release record and version history;
  • exception or limitation rationale.

3. Evidence of machine-readable marking

For synthetic audio, image, video, or text, a reviewer may need to examine the marking mechanism, representative outputs, detectability testing, and known limitations. A policy statement that content “will be watermarked” does not show which mechanism operated or whether it remained effective across output formats.

Representative artifacts

  • marking or provenance specification;
  • configuration and release record;
  • representative marked outputs;
  • detector or verification test results;
  • format-conversion and degradation testing;
  • documented technical limitations and fallback measures.

4. Evidence of deepfake and public-interest disclosure

Where a deployer publishes deepfakes or certain AI-generated or manipulated text on matters of public interest, the package should connect the disclosure to the publication workflow. Editorial review should be supported by an attributable record rather than asserted after the fact.

Representative artifacts

  • publication and labelling standard;
  • content classification decision;
  • human review or editorial responsibility record;
  • approver identity and publication timestamp;
  • the final published version and disclosure placement;
  • change or correction history.

5. Evidence that the measure was tested and maintained

Reviewers will often need to distinguish a one-time design decision from an operating process. That requires test protocols, acceptance criteria, reviewer conclusions, and evidence of what changed when the system or policy changed.

A strong package preserves the relationship between the requirement, supplied artifact, verification action, system version, and remediation decision. Candidate evidence relationships may be accelerated through technical matching, but the relationship included in the package should be human verified.

The five outputs of an Article 50 evidence package

  1. Evidence Inventory: notices, marking specifications, test records, representative outputs, approvals, and limitations.
  2. Mapping Summary: human-verified relationships between supplied artifacts and the defined Article 50 review elements.
  3. Gap Register: missing, weak, stale, unattributed, inaccessible, or contradictory evidence.
  4. Ordered Remediation Plan: the sequence required to improve reviewability.
  5. Chain-of-Custody Statement: artifact source, owner, collection date, version, hash, verification action, and ledger range.

What the evidence package does not prove

Evidence mapping does not by itself establish legal compliance, control effectiveness, or continuous operation. It establishes what was supplied, what was verified, how it relates to the defined review context, and what remains unresolved at the assessment date.

Official sources

ARTICLE 50 EVIDENCE READINESS

Turn transparency measures into a reviewer-ready evidence package.

The $3,500 fixed-scope ATO Readiness Baseline covers one defined AI system and review context. No annual license is required.

Request Article 50 Baseline Intake