Marking AI-generated content: the December 2, 2026 obligation explained

Two AI Act obligations look alike but target different people: technical marking of generated content, due December 2, 2026 for systems already on the market, and disclosure of AI-written text, applicable since August. Here is what applies to your project.

The obligation to mark AI-generated content has a December 2, 2026 deadline: providers of systems generating text, sound, or images placed on the market before August 2, 2026 must comply with Article 50(2) of the AI Act by that date, according to the Future of Life Institute's timeline.

Two obligations that look alike

The first falls on the generative system's provider and concerns a technical, machine-readable marking embedded in the output. This is Article 50(2), which carries the December deadline for systems predating August. The second falls on whoever deploys the system and publishes the result: text produced with AI and published to inform the public on a matter of public interest must be identified as such. Confusing them leads people to believe they comply because their tool does, or to add a disclosure where the text does not require one. The first useful step is therefore to identify your position in the chain for every piece of content you publish.

What already applies

AI Act transparency obligations have applied since August 2, 2026, Toute l'Europe notes in its article updated August 3, 2026: chatbot users must be told they are talking to AI unless it is obvious, generated content must carry a marking, deepfakes a visible disclosure, and AI-generated text of public interest must be identified. The same source gives the penalties: €35 million or 7% of worldwide turnover for general failures, €15 million or 3% for transparency failures. The most demanding part, concerning standalone high-risk systems, was postponed to December 2, 2027, giving recruitment or credit software more time but granting none for other obligations.

The exemption everyone is discussing

An opinion piece by Harold Paris (ELOQIO), published by Journal du Net on September 1, 2026, argues for a specific reading of Article 50 paragraph 4: the disclosure requirement for AI-written text would fall away once a person competent in the subject has reviewed it and editorial responsibility is stated in the legal notice. This is a practitioner's reading, not an authoritative position, and should be discussed with your adviser before basing publishing practices on it. It nevertheless overlaps with what the CNIL's note on agentic AI and case law emphasise elsewhere: a meaningful review can change the text.

The number we are not adopting

The same opinion piece calculates a €60,000 ceiling for a company with €2 million turnover, applying 3%. The calculation is consistent with published scales, comes from no authority, and no penalty of that amount has been imposed to date. We cite it for what it is, an approximate figure put forward by the author of an opinion piece, and do not present it as official. Articles repeating it without that qualification sell fear rather than information.

Who is the provider in your case?

The question becomes concrete for a custom application calling a market model to draft responses, summarise files, or create visuals. You did not build the model, but you make a generating system available. The division of roles between the model provider and you depends on what your product does and how it is offered: to employees, customers, or the public. This belongs in the specification rather than being discovered in December, alongside your software's accountability for its vulnerabilities. Have your adviser determine the classification, and keep their written answer with the rest of the file.

In practice, before December

Open your product and inventory everything it creates without a human writing it: automatic messages, summaries, descriptions, catalogue images. For each line, note two things: whether it goes outside the organisation and whether someone reviews it first. Items published without review need a decision: a disclosure or an organised review with a named person responsible. In Maestro, every agent-produced document waits for your agreement before the next stage starts, and every version can be found if a decision must be undone. These mechanisms support your review practices; they do not make any product compliant, and nobody should sell you that claim.