Guides · 2026-10-11
EU AI Act Article 50: Transparency Rules for Publishers
Since 2 August 2026, publishing AI-generated content in the EU without proper disclosure isn't just sloppy — it's a finable offence. Article 50 of the EU AI Act makes transparency a legal duty, with fines up to €15 million or 3% of worldwide annual turnover. Here's what the rule actually requires, who it catches, and how publishers comply in practice.
What Article 50 actually says
Article 50 sits in the AI Act's transparency chapter. It doesn't ban AI-generated content; it bans undisclosed AI-generated content. The core duty: providers and deployers of AI systems that generate synthetic audio, image, video or text content must ensure that content is marked in a machine-readable format and that recipients are informed it is AI-generated — typically through a visible label or disclosure.
Two obligations, two audiences. The machine-readable marking (metadata, watermarks, cryptographic manifests) exists so platforms, regulators and other automated systems can detect synthetic content at scale. The visible disclosure exists so the human on the other end isn't fooled. You need both.
Provider vs. deployer: who owes what
The Act splits duties. Providers — the companies building the AI systems — must build technical solutions that make marking possible: watermarking capabilities, metadata outputs, disclosure tooling shipped with the model. Deployers — that's you, the marketing team, agency or creator publishing AI-assisted ads and content — must actually apply the markings and disclosures to the content you put out.
This is the part most teams miss. Buying a compliant tool doesn't make your output compliant; the duty to label what you publish sits with the publisher. If your agency generates ad creatives with an AI image tool and ships them without labels, the enforcement target is the agency, not the model vendor.
Deepfakes and the higher bar
Article 50(4) raises the bar for deepfakes: AI-generated or manipulated image, audio or video content that depicts real persons, or that could mislead about events, must be disclosed prominently. A buried hashtag doesn't cut it. For election-related or public-interest content, several member states have layered on additional national rules — check local guidance wherever you publish.
Deadlines that already passed
The transparency obligations became applicable on 2 August 2026. A grace arrangement for certain legacy systems ran to 2 December 2026 — that window is now closed or closing, depending on when you read this. Enforcement is live, and national market-surveillance authorities are the ones holding the fine pen.
What compliance looks like in practice
A defensible Article 50 posture has four layers:
- Embedded provenance: preserve or attach machine-readable markings — C2PA content credentials where your tooling supports them, XMP or EXIF fields at minimum. Never strip metadata your generator added.
- Visible disclosure: a clear label a viewer can't miss — "AI-generated", "Created with AI", or the platform's native disclosure tag — placed at first exposure, not three clicks deep.
- Per-asset records: for every published asset, keep a record of what was generated, what was marked, and where it was published. Regulators ask for evidence, not intentions.
- Drift monitoring: social platforms recompress uploads and routinely strip metadata. Re-check published assets: if the label you shipped didn't survive publishing, your compliance didn't either.
The fine math
Article 50 breaches fall under the Act's mid-tier penalties: up to €15 million or 3% of worldwide annual turnover, whichever is higher. For a mid-size agency, the turnover limb is usually the scary one. And fines compound with platform penalties — Google Ads can suspend accounts for undisclosed synthetic content independently of any regulator.
Bottom line
Article 50 is a paperwork-and-labels law, not a technology ban. Complying is mostly operational discipline: mark everything, label visibly, keep records, and re-verify after publishing. The teams that get fined will be the ones who treated disclosure as someone else's problem. DisclosureGuard automates the checking part — deterministic provenance reads against versioned Article 50 rules, with a 0–100 score per asset — but the duty itself is yours. Start with your highest-reach assets and work down.
Informational explainer, not legal advice. Check a specific asset against the current rules with a free DisclosureGuard scan.