AI-transparency obligation testing (EU AI Act Article 50)

How it works

Article 50 of the EU AI Act sets transparency obligations that apply from 2 August 2026, and each is a testable requirement. Providers must design systems that interact directly with people so that those people are informed they are interacting with an AI system, unless that is obvious to a reasonably well-informed, observant and circumspect person. Providers of systems that generate synthetic audio, image, video or text must ensure the outputs are marked in a machine-readable format and detectable as artificially generated or manipulated, as far as technically feasible given the type of content, cost and the state of the art, with exceptions such as assistive editing functions. Deployers have their own duties, including informing people exposed to emotion-recognition or biometric-categorisation systems, and disclosing deepfakes and AI-generated text published to inform the public on matters of public interest, unless that text has undergone human review or editorial control and someone holds editorial responsibility. Law-enforcement uses have separate exceptions. And the information must be given clearly and distinguishably, at the latest at the time of first interaction or exposure, and meet accessibility requirements. Testing turns each obligation into a check. Disclosure: in every channel the system reaches — web chat, voice, email, a messaging integration, an embedded widget — is the disclosure present before or at the first exchange, visible, and accessible to a screen reader. Marking: does every generated image, audio file or document carry the machine-readable mark, and does the mark survive the transformations the system's own pipeline applies — resizing, format conversion, export — to the extent the marking technique is designed to. Persistence, a design choice beyond the Article's first-interaction minimum: does the system keep to the disclosure — never denying it is an AI when asked — under persona instructions, user requests to drop it, and long conversations. Coverage across channels is the usual gap, since the disclosure gets designed for the main interface and forgotten in the integration added later.

When to use it

Any AI system offered in the EU that talks to people or generates content; before launching in a new channel or integration; after changes to the persona, the output pipeline or the media formats the system produces; and as part of a compliance review, to turn a legal reading into evidence that the system behaves as the reading says it should.

Limitations

This tests whether the system does what a legal interpretation says, not whether the interpretation is right — scope, exemptions and 'obvious from the context' are legal questions for counsel, and the test can only encode the answer. Machine-readable marking depends on technical standards and detection tools that are still settling, so a mark that passes today's detector may not satisfy tomorrow's guidance. Amendments and implementing guidance change what is required and when — Regulation (EU) 2026/1744 (the Digital Omnibus on AI) gave providers of generative systems already on the market before 2 August 2026 until 2 December 2026 to meet the Article 50(2) marking duty — so the requirements behind the suite need re-checking against the current consolidated text. And Article 50 is one part of the Act; a system can pass every check here and still be out of compliance with obligations that depend on its risk classification.

Cite this