Are your Article 50 transparency obligations for AI content in order: evidence for an enforcement query
A market surveillance authority asks how a product that generates or publishes AI content meets Art. 50 of Regulation (EU) 2024/1689, and the provider answers with test records and architecture documents, since a policy statement records no observation. Art. 50(1) and 50(2) are provider duties and Art. 50(4) is a deployer duty. Art. 74(1) gives the authority the document and sample powers of Regulation (EU) 2019/1020. Art. 50 has applied since the general application date In force. This page gives the assembly sequence, the evidence pack and the consequence of an empty reply.
Where this starts
The query names a product, a period and one or more Art. 50 paragraphs, and it asks for the material that shows how the duty is met. Its legal basis is Art. 74(1) of Regulation (EU) 2024/1689, which applies Regulation (EU) 2019/1020 to AI systems covered by the Act; Art. 14(4)(a) of that Regulation empowers the authority to require documents, technical specifications, data and information on conformity, and Art. 14(4)(j) lets it obtain product samples, including under a cover identity, and inspect them. A query can also follow a complaint under Art. 85, which any natural or legal person may submit to the market surveillance authority. The authority can therefore generate outputs as an ordinary user, take a screenshot, run a detector and compare the result with what the reply claims. The generative AI hub carries the obligation set with its two dates; Regulation (EU) 2026/1744 made, in the words of the regulation entry, no change to the Art. 50 date, and it added a grace period for the Art. 50(2) marking duty on systems already on the market Deferred, so a legacy system answers with its grace-period record as well.
Each paragraph asks for a different kind of evidence, and the Art. 50 obligation record splits them into testable, documentation and process elements. For Art. 50(1) the evidence is an observation: the notice as a new user sees it at the first interaction, its location in the interface, and the recorded assessment where the provider relies on the exemption for an interaction that is obvious to a reasonably well-informed, observant and circumspect person. For Art. 50(2) the evidence is a set: the marking architecture description per output modality, the justification of the technical solution against the four criteria of effectiveness, interoperability, robustness and reliability, the presence and survivability test results, the detection solution a party other than the provider can run, and the detection and false-positive rates against a target the provider justified. Art. 50 sets no numeric threshold, no harmonised standard is cited in the Official Journal for it and no JTC 21 draft covers it, so no presumption of conformity is available for any technique; the level of the source hierarchy in use is the published Code of Practice on Transparency of AI-Generated ContentPublished and the published Commission guidelines on Article 50 transparency obligationsPublished, with the published C2PA Content Credentials specificationPublished as the technique reference for the metadata and provenance layer. For Art. 50(4) the evidence is the label on each distribution surface and, where the editorial-control exemption is relied on for public-interest text, the record naming the person who holds editorial responsibility and the review measures. Where the chatbot notice itself has never been checked, the chatbot scenario covers that check; this page assumes the product is live and the question is what to hand over.
The layer set is the element where the published sources set the most specific expectation. Measure 1.1 of the published Code expects at least two machine-readable layers for audio, image and video, signed metadata and an imperceptible watermark, each with a detection mechanism, and the guidelines expect a provider that has not signed the Code to record a gap analysis against it. Which technique set counts as adequate under Art. 50(2) is an unsettled point; the pack records the layer set per modality, the reading applied and the alternative reading next to the result.
The sequence
The work runs in the order below: the query record first, then the role and paragraph assignment, then the frozen documents, then the tests on the versions the query names, then the pack and its gap list.
Record the authority, the legal basis cited, the product versions and paragraphs named and the reply date.
Record Request record: authority, legal basis cited, product and versions named, paragraphs named, period, reply date set
[Regulation (EU) 2024/1689, Art. 74(1); Regulation (EU) 2019/1020, Art. 14(4)(a)]
Determine the role for each named function and assign the function to the Art. 50 paragraph that applies.
Record Role determination record; Article 50 applicability record, one row per function
[Regulation (EU) 2024/1689, Art. 3(3), 3(4) and 50(1) to (4); PR-CLS steps 8 to 9 and 35]
Observe each claimed exemption and record whether the system was on the market before the Art. 50 date.
Record Article 50 exemption observation record; Art. 111(4) legacy status record
[Regulation (EU) 2024/1689, Art. 50(1), 50(2) and 111(4); PR-CLS steps 36 to 37]
Freeze the marking architecture description, the detection access terms and the target justification with a hash per document.
Record architecture-description, detection-access-record and target-justification-record, each with version and content hash
[Regulation (EU) 2024/1689, Art. 50(2); PR-50-MARK steps 4 to 6]
Generate samples on each output path of each named version and run presence, survivability and independent-reader checks.
Record manifest-validation-log, watermark-detection-log, survival-matrix and independent-reader-log per named version
[PR-50-MARK steps 11 to 16 and 19 to 24]
Run the detection solution from outside the provider's environment and compare the rates with the target.
Record external-detection-log, control-detection-log, detection-rate-summary and target-comparison
[Regulation (EU) 2024/1689, Art. 50(2); PR-50-MARK steps 17, 18, 25 and 36 to 37]
Compare the layer set per modality with the published Code and record the reading applied.
Record layer-set-comparison per modality, with the reading applied and the alternative reading noted
[Code of Practice on Transparency of AI-Generated Content, Section 1, Measure 1.1; PR-50-MARK step 38]
Observe each notice, label and disclosure on each surface as a new user and classify each one.
Record interaction-notice-log, deepfake-label-log, label-persistence-log, text-disclosure-log and disclosure-classification per surface
[Regulation (EU) 2024/1689, Art. 50(1), 50(4) and 50(5); PR-50-MARK steps 26 to 27, 29 to 33 and 40]
Assemble the reports, records and run log into an indexed pack and list each gap with its re-test criterion.
Record Evidence pack index: item, version, hash, Art. 50 paragraph, status; remediation list with a re-test criterion per gap; run-log
[PR-50-MARK steps 41 to 45]
Evidence this situation ends with
| Artifact | Answers | Produced by |
|---|---|---|
| Article 50 applicability and exemption determination record (PR-CLS output) | Art. 50(1) to (4): the paragraph assigned to each function, the role held, and the observed behaviour against each claimed exemption | PR-CLS v1.0 |
| Marking architecture description per output modality | Art. 50(2): the metadata, watermark and provenance layers in use, the pipeline point where each is applied, and the detector that reads each layer | the provider |
| Technical solution justification against the Art. 50(2) criteria | Art. 50(2): why the chosen layers are effective, interoperable, robust and reliable, with the feasibility limits, the content-type specificities and the state of the art relied on | the provider |
| Marking presence and survivability test report (PR-50-MARK output) | Art. 50(2), robustness: the mark present per sampled output and path, the survival matrix per transformation and layer, and the layer set compared with the published Code | PR-50-MARK v1.0 |
| Third-party detectability verification report (PR-50-MARK output) | Art. 50(2), detectable, interoperable and reliable: the detection run under third-party access terms, the independent validator agreement, and the detection and false-positive rates against the justified target | PR-50-MARK v1.0 |
| Interaction disclosure and label audit record (PR-50-MARK output) | Art. 50(1), 50(4) and 50(5): each notice, label and disclosure observed on each surface, with its moment, location and classification | PR-50-MARK v1.0 |
| Code of Practice adherence or alternative-means record (Art. 50(7)) | Art. 50(7): adherence to the published Code, or the gap analysis against it and the alternative means used to show conformity with Art. 50(2) and 50(4) | the provider |
| Remediation list (PR-50-MARK output) | Each gap with its Art. 50 paragraph, the observed behaviour, the record reference and the re-test criterion that closes it | PR-50-MARK v1.0 |
Each artifact follows a specification with a synthetic sample on the deliverables page: the applicability record and the audit record follow the assessment memo specification, the two test reports follow the test report specification, the indexed set follows the evidence pack specification, and the remediation list follows the findings register specification. Every report carries the run log with the protocol version, the tool versions and the sample hashes, so the authority can repeat a check on the same sample.
If nothing is done
Art. 79(2) of Regulation (EU) 2024/1689 requires a market surveillance authority that has sufficient reason to consider that a system presents a risk to evaluate it against all the requirements and obligations of the Regulation and, where the system does not meet them, to require the operator to take corrective action, withdraw the system or recall it within a period the authority may prescribe, no longer than the shorter of fifteen working days or the period in the relevant Union harmonisation legislation; Art. 79(5) requires provisional measures on the national market where the operator does not act within that period. A breach of the transparency obligations of Art. 50 sits in the penalty tier of Art. 99(4)(g), an incorrect, incomplete or misleading reply to the authority's request sits in the tier of Art. 99(5), and Art. 99(7)(f) counts the degree of cooperation with the authority when the amount is set. The action on the provider's side is to answer by the reply date with the frozen documents, the test reports and the indexed pack, and to state each gap the tests found. The corrective action is the one the gap names: a second layer added where metadata was the only mark and a screenshot removed it, a detector opened to parties other than the provider, or a notice moved into the interface at the first interaction. Each correction is followed by the re-test that the process element of the obligation record requires after a change to the model, the encoder, the surface or the marking library.
Services, guides and tools for this situation
The Article 50 transparency and marking/watermark testing service runs PR-50-MARK v1.0 on the versions the query names and assembles the two reports and the audit record, within the limits stated on what we cannot do; the role and paragraph assignment that opens the sequence is a PR-CLS v1.0 output of the classification and readiness assessment.
- Obligations
- Guides
- ICP hubs
Every inquiry receives a reply within 24 hours. Scope and availability are discussed per request.