Transparency claims need testable evidence
Article 50 transparency obligations address several different situations: informing people when they interact with certain AI systems, marking generated or manipulated outputs in machine-readable form, and labelling deepfakes and certain public-interest text.
The Commission's transparency code separates provider measures for marking and detection from deployer measures for visible labelling. Adherence to the code is voluntary, while applicable Article 50 obligations are legal requirements.
Procurement teams should therefore ask which role the supplier performs, which service features are in scope and what evidence shows the measures work in the buyer's deployment context.
1. Scope and responsibility statement
Request a feature-level map showing where the service generates or manipulates text, images, audio or video; where users interact directly with AI; and which party controls publication or presentation.
The statement should identify provider, deployer and upstream responsibilities, relevant exceptions or human-review processes, supported channels and geographic availability. It should be approved by a named owner and tied to the current product version.
2. User-interaction disclosure evidence
For chatbots and other interactive systems, request screenshots and test records showing when and how people are informed that they are interacting with AI. Cover first use, returning sessions, embedded experiences, voice channels and accessibility modes.
Evidence should show that the disclosure is timely, clear and not hidden behind optional help content. Test localisation, responsive layouts and integrations where the buyer controls part of the interface.
3. Machine-readable marking specification
Ask providers of generative systems to document the technical method used to mark outputs, supported media types, persistence through common transformations, interoperability, known limitations and monitoring.
Request representative marked outputs and independent or internal test results. The Commission describes effective, interoperable, robust and reliable marking as far as technically feasible, taking account of content type, implementation cost and the state of the art.
A product roadmap is not current evidence. Record which formats and delivery paths are supported today and what happens when marking cannot be applied.
4. Visible labels and editorial-control records
For deepfakes and relevant AI-generated or manipulated public-interest text, request the label design, placement rules, trigger logic, user guidance and examples across channels.
Where a supplier relies on human review and editorial responsibility, ask for the workflow, accountable role, review criteria, approval record and audit trail. A checkbox without an operating process is not persuasive evidence.
The EU has published icons that deployers may use, but the buyer should still test whether the full disclosure is understandable in context.
5. Assurance, incidents and change control
Request the test plan, coverage, failure thresholds, monitoring metrics, incident route and corrective-action records for transparency controls. Include examples of detected failures and how the supplier verified remediation.
Connect transparency evidence to the supplier change log. A new model, output format, user interface, integration or content-processing step may affect whether disclosures and marks remain effective.
Require notice of material changes and refreshed evidence. The pack should state its version, applicable product release, owner and review date.
Turn the pack into an acceptance test
Before approval, procurement, product and compliance teams should test representative user journeys and sample outputs. Record expected disclosure, observed result, evidence reference, exception and owner for remediation.
Repeat the test after material changes and on a risk-based schedule. For wider context, read our Article 50 business checklist and supplier change-log guide. AI Act Ready helps buyers turn transparency promises into reviewable acceptance evidence.