Who must label AI-generated content, and from when?

Synthetic content disclosure is the duty, set out in Article 50(2) of Regulation (EU) 2024/1689 (the AI Act), requiring providers of AI systems, including general-purpose AI systems, that generate synthetic audio, image, video or text to ensure the outputs are marked in a machine-readable format and detectable as artificially generated or manipulated. The text asks for technical solutions that are "effective, interoperable, robust and reliable as far as this is technically feasible", taking into account the specificities of the content type, the costs of implementation and the generally acknowledged state of the art. It is not an absolute quality standard for the marking, it is a duty of means calibrated to what is realistically achievable today for that type of output. The duty applies from August 2, 2026, but with a qualification on the actual deadline, explained further below, that concerns only systems already on the market before that date.

Who is obligated, and what this obligation is not

The duty falls on the provider, the party that develops or commissions the system generating the synthetic content, not on the deployer using it in its own business. It must be kept separate from Article 50(1), which imposes a different, general transparency obligation when a person interacts with an AI system (a chatbot that must disclose it is one): that provision concerns the interaction, this one concerns the output produced and distributed. Confusing the two means believing that a banner saying "you are talking to a bot" also satisfies the duty to mark the generated image or video in a machine-readable format, which it does not.

Available technical tools

The rule does not mandate a specific standard, and in practice two families of solutions cover most cases: watermarking, a statistical signal embedded in the content itself, and provenance standards based on signed metadata, of which C2PA is the most widely adopted industry example for images, audio and video. Neither solution is tamper-proof (metadata can be stripped by recompression, text watermarking is fragile on short content), so a provider that has to demonstrate compliance in practice combines more than one signal rather than relying on just one. Anthropic made this dual approach public for Claude on August 11, 2026: an invisible watermark embedded in text that travels with the content even after copy-paste, and signed C2PA provenance metadata for generated files (PNG, JPG, SVG), with an explicit caveat that a detected signal indicates provenance, not absolute proof of authorship (source: The Register, August 11, 2026).

An enterprise example

An Italian marketing agency produces an advertising video with a synthetic voice generated by a text-to-speech tool and a product image created with a generative model for a small-business client, and publishes both on August 5, 2026 with no label or provenance metadata. If the tool in use counts as a provider under the regulation, that output falls under the Article 50(2) duty with no transitional period at all, because the system was placed on the market after August 2, 2026: the omission is not a procedural detail, it is the violation itself. That case differs from, but is not unrelated to, one where synthetic content is used for outright fraud: when marking is missing on a deepfake used to impersonate a real person, the missing label makes the harm worse, because it removes the one signal that could have let the recipient suspect an artificial origin before acting.

Why it matters for decision makers

The August 2, 2026 deadline is not a single date for everyone: the Digital Omnibus on AI, now in force, introduced a four-month transitional period, until December 2, 2026, but only for AI systems already placed on the market before August 2, 2026. For systems placed on the market on or after August 2, 2026, the obligation is immediate and carries no extension at all. Anyone commissioning or providing AI-generated content therefore needs to know which system they are using and since when, not just remember "the deadline is in August": a provider that launched its tool in July 2026 has until December to mark that system's outputs, one launching it in September has no grace period at all.

This content is for informational purposes only and does not constitute legal advice: to apply it to your specific case, in particular to establish whether your role is that of provider or deployer, involve your legal counsel or your DPO.