EU AI Act Article 50 Explained
Article 50 of Regulation (EU) 2024/1689 sets transparency obligations for particular AI systems and actors. It is not a blanket requirement for every website that mentions AI. Start by identifying the system, its intended use, who is the provider or deployer, and whether an exception might apply.
When does it apply?
Article 50 applies from 2 August 2026. Its requirements cover four different situations:
- Systems intended to interact directly with natural persons.
- Systems generating synthetic audio, image, video, or text.
- Emotion-recognition and biometric-categorisation systems.
- Deepfakes and certain AI-generated text published to inform the public on matters of public interest.
Direct interaction — Article 50(1)
Providers must design systems intended to interact directly with people so that people are informed they are interacting with AI, unless that is obvious in context to a reasonably well-informed, observant, and circumspect person. The information must be clear, distinguishable, accessible, and provided by the first interaction.
For a chatbot, a plain-language notice in or immediately before the chat is the safest pattern. A generic statement such as “AI-powered” may be relevant evidence, but it is not automatically enough on its own.
Synthetic output marking — Article 50(2)
Providers of in-scope systems generating synthetic audio, image, video, or text must make outputs machine-readable and detectable as artificially generated or manipulated. The technical solution should be effective, interoperable, robust, and reliable as far as technically feasible.
C2PA/Content Credentials can be a strong provenance method, but Article 50 does not mandate C2PA. Likewise, a visible label or HTML metadata tag does not by itself prove that output-level machine-readable marking is in place. Ordinary assistive editing that does not substantially alter the input or its semantics is outside this obligation.
Biometric, deepfake, and public-interest content
Deployers of emotion-recognition or biometric-categorisation systems must inform people exposed to those systems. Deployers must also disclose deepfake image, audio, or video content. AI-generated or manipulated text published to inform the public on matters of public interest must be disclosed unless it has undergone human review or editorial control and a natural or legal person holds editorial responsibility.
These cases are contextual. A page crawl cannot reliably determine the responsible actor, intended purpose, editorial review, or statutory exception.
What an automated screening can establish
An automated public-page screening can capture visible disclosures, metadata, and other observable evidence. It cannot establish legal compliance by itself. Treat its findings as:
- Potential gaps found — evidence supports a focused follow-up.
- No observable gaps — relevant public evidence exists, but limitations remain.
- Needs applicability review — the crawl cannot prove Article 50 scope or responsibility.
The SentinelGo score is an evidence-readiness score, not a legal compliance score.
Penalties
Article 50 falls within Article 99(4): the statutory maximum is up to EUR 15 million or 3% of total worldwide annual turnover, whichever is higher. This is not an automatic or predicted fine; authorities assess individual cases and Article 99 includes SME/start-up proportionality rules.
Build an evidence trail
Document the system inventory, intended purpose, roles, user journey, exact disclosure wording and timing, content-marking implementation, tests, exception analysis, and human editorial controls. That record is usually more valuable than a single “compliance” badge.
For the official legal text and current Commission guidance, see EUR-Lex and the European Commission’s Article 50 guidance.