Article 50 transparency obligations, explained

Do I have to tell users they're interacting with AI?

Yes, if your system is intended to interact directly with people — Article 50(1) puts the duty on the provider: design the system so users know they're dealing with AI, at the latest at first interaction, in a clear and distinguishable way. The Commission's July 2026 transparency guidelines read "interacting" broadly — a single prompt followed by a single reply counts — while carving out things like spam filters and pure backend decision support. There is an exception when the AI nature is "obvious" to a reasonably observant person, but the guidelines interpret it restrictively: it works for developer tools aimed at specialists, not for general-public consumer products. The practical answer for a small product: add the one-line label. It costs nothing and moots the gray zone.

Do I have to mark AI-generated content machine-readably?

Article 50(2) requires providers of systems that generate synthetic audio, image, video or textto mark outputs in a machine-readable format, detectable as AI-generated — "as far as technically feasible." The Code of Practice on transparency of AI-generated content (final version June 2026) is the operative benchmark: for files that can carry metadata, signed metadata plus a watermark; for free-form text, an imperceptible watermark for longer output, with very short text excepted. The feasibility qualifier matters for small products: the guidelines state you are not obliged to use technology that isn't available on the market. If you stream text from a model API that offers no watermarking, document that assessment in writing, disclose visibly at the point of output, and adopt upstream marking the moment your model provider ships it. How authorities will treat downstream API providers on this point is genuinely unsettled — which is exactly why the written feasibility assessment is worth having. For concrete marking options across text, images, audio and video, see how to mark AI content.

Does the assistive-editing exception cover paraphrasing and rewriting?

No. Article 50(2) exempts systems with an "assistive function for standard editing" that don't substantially alter the input. The guidelines' examples draw the line clearly: grammar and spell-checking sit inside the exception; AI summaries and paraphrasing or rewriting that changes style, structure or meaning sit outside it. A rewriter whose entire point is changing the style of a text alters the input in exactly the dimension the guidelines name. If your product rewrites, the marking obligation applies to it.

What if I publish AI-generated text on my site or blog?

Article 50(4) binds deployers— whoever publishes, not whoever built the tool. It requires disclosure when AI-generated text is published to inform the public on matters of public interest, with an exception where the text went through human review and someone holds editorial responsibility. Three things follow. If you run a tool, you don't police your users' publications. If you publish AI text yourself on such matters, disclose or edit it under editorial responsibility. And purely personal, non-professional publishing is outside the Act entirely.

When do these obligations apply?

Article 50 has applied since 2 August 2026. One transitional rule softens the marking duty: systems already on the market before that date have until 2 December 2026 to comply with the 50(2) marking requirement — the disclosure duties applied immediately. Content generated before August 2026 never needs retroactive labeling. The full picture, including the AI Omnibus changes, is on our deadlines page.

What happens if I ignore Article 50?

Non-compliance carries fines up to €15 million or 3% of worldwide turnover — for SMEs, whichever is lower, and national authorities must weigh proportionality: nature, gravity, duration, cooperation. Since 2 August 2026 anyone can also lodge a complaint with a market surveillance authority. In practice, a small product with visible disclosures, a marker on generated output, and a written feasibility file is a very unattractive enforcement target. Run the self-check to see which paragraphs apply to your product, or read how we implemented all of this ourselves.