Article 50 is the disclosure article — what your AI product must tell the people interacting with it, and what synthetic content must be labelled. It reaches a much wider set of organisations than the high-risk regime, and its watermarking grace period is one the Digital Omnibus would compress. This is the operational read for product, engineering, and compliance leads — built to amplify your in-house expertise.
What Article 50 requires
Article 50 of Regulation (EU) 2024/1689 imposes transparency obligations on providers and deployers of certain AI systems, regardless of high-risk status. Four duties matter most:
- Chatbot disclosure (Art. 50(1)) — providers of AI systems interacting directly with natural persons must design them so the person knows they are interacting with AI, unless obvious in context.
- Synthetic content marking (Art. 50(2)) — providers of AI systems generating synthetic audio, image, video, or text must mark output as artificially generated in a machine-readable format, using solutions that are effective, interoperable, robust, and reliable as far as technically feasible.
- Deepfake disclosure (Art. 50(4)) — deployers of AI that generates or manipulates image, audio, or video deepfakes must disclose the content has been artificially generated or manipulated.
- AI-generated text disclosure (Art. 50(4)) — deployers of AI generating or manipulating text published to inform the public on matters of public interest must disclose it — with carve-outs.
Article 50 applies on the same day as the Act's main application — currently 2 August 2026 under Regulation 2024/1689. See the European Commission AI Act timeline for the full sequence.
The grace-period compression under the Omnibus
The original Regulation gives providers and deployers a six-month grace period after the Article 50 application date to bring synthetic-content marking solutions into compliance. Under the Digital Omnibus on AI — provisionally agreed by the Council and Parliament on 7 May 2026 (PRESS 299/26) — that grace period would be compressed from six months to three months, with a new deadline of 2 December 2026.
The Omnibus deal is provisional. Until formally adopted, the original six-month window remains the law. But for product planning purposes, a three-month grace window is the prudent assumption — starting the engineering work in October 2026 leaves no contingency if the Omnibus is adopted as agreed.
Two practical reads. First, the obligation itself is unchanged; what shortens is the time you have to wind up the marking solution after Article 50 begins to apply. Second, the marking obligation sits with the provider — if you build the GPAI model, the synthetic-content marking is on you, and your deployer customers will expect to inherit a compliant output stream.
What to build in your product to comply
Article 50 is not legal-team work — it is product-engineering work that legal scopes. Three workstreams cover most teams:
- Chatbot disclosure surfaces. Every interaction surface where a natural person talks to an AI system needs the disclosure. That means the chat input, the voice agent opening, the in-app assistant entry point. The disclosure can be short — "You're talking to an AI assistant" — but it has to be there, and it cannot be buried in terms.
- Synthetic content marking pipelines. Article 50(2) names machine-readable marking. Practically that means watermarking for image and video, provenance metadata (such as C2PA Content Credentials) for image, video, and audio, and either watermarking or provenance for text where technically feasible. Track the European Commission policy pages for the harmonised standards as they land.
- Deployer-facing disclosure controls. Deployers under Art. 50(4) need product affordances to make the deepfake or AI-text disclosure they are required to provide. If your product is the tool a deployer uses, expose a clear, configurable disclosure setting — do not make them build it themselves.
Sequence the work: chatbot disclosure first (cheapest), provenance metadata second (infrastructure already exists), watermarking third (heaviest lift). All three need to be live before the grace period closes — whichever date that is.
Edge cases — public-interest text, satire, journalistic exceptions
Article 50(4) catches teams off-guard because of built-in exceptions that need careful reading.
- AI-generated text on matters of public interest must be disclosed — unless the content has been subject to human review or editorial control and a natural or legal person holds editorial responsibility. Many newsrooms qualify; many programmatic-content publishers do not.
- Deepfakes must be disclosed — but the obligation can be tailored where the content is part of an evidently artistic, creative, satirical, or fictional work. The carve-out is narrower than commentary suggests; "obviously fake" and "evidently artistic" are different tests.
- Law-enforcement uses carry narrow exceptions for both deepfake and AI-text disclosures where authorised by law.
If your product sits in a grey area, document the analysis and editorial-control evidence — that is what a regulator will ask for.
Where to read the source material
- Regulation (EU) 2024/1689 — full text on EUR-Lex — Article 50 sits in Chapter IV.
- Council press release PRESS 299/26 — 7 May 2026 Omnibus agreement — for the provisional grace-period compression.
- European Commission AI Act policy page — for harmonised standards and the timeline.
Generate your data processing agreement → /tools/dpa-generator
Or generate your AI policy → /tools/policy-generator
Qualified review still required. Outputs are AI-generated starting-point documents — not a substitute for qualified legal or compliance advice.