Machine-readable is not the same as noticeable

A provenance marker and a visible disclosure solve different problems. The marker lets a platform, investigator, or downstream system detect that content was generated or manipulated. Notice helps a person understand what they are encountering at the moment that understanding matters. One can survive without the other. A file can carry excellent metadata while a screenshot, transcription, or copied passage reaches people with no visible context.

The European Commission's July 20 transparency guidelines address Article 50 obligations that begin applying on August 2, 2026. The underlying AI Act distinguishes machine-readable marking of synthetic outputs from disclosure duties that arise in particular human-facing uses. Operational programs should preserve that distinction instead of treating one label as a universal transparency control.

Machine readability is a claim about an interface between systems. Noticeability is a claim about a person's experience. They have different failure modes, test populations, owners, and evidence. Combining them in one compliance checkbox makes it difficult to see which protection failed when synthetic content travels.

Design two linked controls

The technical control should specify where the marker is inserted, which formats and modalities are supported, how it survives transformations, and how a relying party can detect it. The notice control should specify who must see what, in which language and presentation, before or during which exposure. A shared content identifier can connect the two receipts without forcing them into the same mechanism.

User research belongs beside format conformance. A symbol can be technically present and practically invisible. A statement can be readable but ambiguous about whether the whole artifact, one element, or only a translation was generated. Test comprehension in the actual feed, player, document, or service where exposure occurs, including small screens and assistive technologies.

A marker helps a system detect origin; notice helps a person interpret an encounter.

Downstream transformations create the hardest cases. Cropping can remove a visible label. Re-encoding can strip metadata. Speech-to-text can separate words from an audio mark. Quoting can move generated text into a new document. The provider and deployer need an explicit handoff describing which properties are expected to survive and what happens when they do not.

Do not make the marker carry more certainty than the system can establish. It may show that a particular tool processed an artifact without proving that every element was generated, that the content is false, or that no human edited it. The visible explanation should match the actual provenance claim rather than turning a technical signal into a broad judgment about truth.

Metrics should remain separate: detection success across supported transformations, and comprehension or exposure success for the people who receive the content. A single percentage hides whether the system is failing machines, people, or both. The operational review should require both forms of evidence for the contexts that depend on them.

Make the obligation operational

Begin with separate machine-readable marking from human-facing notice while binding both to the same content and release. Express it as a control object rather than a policy summary: scope, triggering condition, applicable system or model version, permitted exception, effective time, evidence source, and the consequence when the control cannot establish compliance. This lets engineering, product, legal, and operations examine the same boundary without pretending their responsibilities are interchangeable.

The minimum receipt should retain marker version and location, content identifier, transformation history, visible disclosure text and placement, applicable exception, and test results. Keep the record proportionate and protect confidential information, but make it possible to determine which rule, artifact, system version, and accountable decision governed the event. A folder of undated screenshots may show that work occurred; it rarely proves that the operative control held for the affected release.

Test the implementation by moving representative text, image, audio, and video through crop, copy, compression, transcription, screenshot, repost, and accessibility paths while checking detection and comprehension. Include ordinary cases, boundary cases, degraded dependencies, and known exceptions. Preserve the starting state, observed output, machine-readable evidence, user-visible result, and any human intervention. Re-run the test after changing a model, content pipeline, interface, standard, provider, or policy interpretation.

The content-provenance owner with the user-experience and legal owners should decide whether the evidence supports continued operation, a narrower scope, a compensating control, or a hold. The owner needs authority over the affected release and access to the evidence. Record unresolved interpretation separately from a technical defect so an engineering patch does not masquerade as a legal conclusion.

Monitor both presence and effectiveness. A marker can exist but be stripped downstream. A disclosure can render but arrive after exposure. A document can be submitted but refer to an obsolete model. Pair a control-presence measure with a consequence or comprehension test, give the claim a review date, and reopen it when a dependency changes.

Maintain a dependency register for the control. Model endpoints, editing pipelines, content formats, user interfaces, identity services, submission portals, vendors, and external standards can change the evidence without changing the policy text. Name which changes invalidate the last test and which monitoring signal proves that the dependency remains inside the reviewed state.

Exercise the exception path as carefully as the ordinary path. Record who can invoke it, which facts they must supply, how long it lasts, what capability or distribution is reduced, and which compensating evidence remains. An exception without expiry and re-entry criteria becomes a second operating model that can silently outlive the reason it was approved.

Keep public and executive claims no broader than the tested boundary. Say which systems, releases, formats, routes, and dates the evidence covers, and identify material exclusions. When a control fails or a dependency moves, update the claim and the remediation record together. A transparent limitation protects more credibility than a universal statement built from a narrow passing test.

— Dispatches · Summit Cognitive