Article 50(4) is short and its scope is narrower than the phrase “AI content must be labelled” suggests. It binds deployers, it covers image, audio and video that is a deep fake, and its first subparagraph contains an exception that is doing more work than most summaries admit.
What Article 50(4) covers
The first subparagraph of Article 50(4) of Regulation (EU) 2024/1689 requires deployers of an AI system that generates or manipulates image, audio or video content constituting a deep fake to disclose that the content has been artificially generated or manipulated. The text is on EUR-Lex.
Four scoping points fall straight out of that sentence:
- Deployers, not providers. The party operating the system in the course of its activity owes the disclosure, not the party that built the generator. A tool vendor does not discharge this for its users.
- Image, audio or video. Text is not in this subparagraph. Text has its own, much narrower rule in the second subparagraph — see the public-interest text labelling page.
- Generated or manipulated. Wholly synthetic content and edits to real footage both count; the provision does not distinguish.
- Constituting a deep fake. Not all synthetic media. A generated abstract image is generated content but is not a deep fake, and the disclosure duty in this subparagraph does not reach it — though Article 50(2) marking may.
Whether a particular piece of content is a deep fake within Article 3(60) and whether a given disclosure is adequate are fact-specific. This page is not legal advice; take advice on your own publication before relying on any reading of the carve-outs.
What counts as a deep fake
The definition is in Article 3(60): AI-generated or manipulated image, audio or video content that resembles existing persons, objects, places, entities or events and would falsely appear to a person to be authentic or truthful.
Two limbs, both required. It must resemble something existing, and it must be capable of falsely appearing authentic. That second limb is why a plainly stylised cartoon of a real politician is a poor fit and a photorealistic one is a clear fit; the definition is about deceptive appearance rather than about subject matter.
The definition is also broader than the popular use of the word in one important respect: it is not restricted to depictions of people. “Objects, places, entities or events” brings in synthetic footage of an event that did not happen and product imagery that misrepresents a real object. A great deal of AI-assisted marketing imagery sits closer to this line than the teams producing it assume.
The artistic and satirical carve-out
The second sentence of the first subparagraph does not remove the duty; it changes its shape. Where the content forms part of an evidently artistic, creative, satirical, fictional or analogous work or programme, the transparency obligations are limited to disclosure of the existence of such generated or manipulated content in an appropriate manner that does not hamper the display or enjoyment of the work.
Read that carefully, because it is routinely reported as an exemption and it is not one. The duty survives; what is relaxed is the manner. A film using a synthetic performance still discloses, but it may do so in a way that does not require a watermark burned across the frame — credits, an accompanying notice, a description at the point of access.
The load-bearing word is evidently. It qualifies the artistic, creative, satirical and fictional categories, and it is an objective test about how the work presents itself rather than a statement of the creator's intention. Satire that is evident as satire is in scope of the relaxation. A fabricated clip of a public figure that its maker characterises as satire but that presents as authentic footage is exactly what the main rule is aimed at, and calling it satire afterwards does not obviously help. Where the line falls between the two is not settled, and it is the kind of question that will be answered by national enforcement practice and eventually by courts rather than by the text.
Two different duties: 50(2) and 50(4)
A single piece of synthetic video can attract two obligations owed by two different parties, and separating them removes most of the confusion in this area.
- Article 50(2) — providers, machine-readable marking. Providers of AI systems, including general-purpose AI systems, generating synthetic audio, image, video or text content must ensure the outputs are marked in a machine-readable format and detectable as artificially generated or manipulated. Solutions must be effective, interoperable, robust and reliable as far as technically feasible. There are exceptions where the system performs an assistive function for standard editing or does not substantially alter the deployer's input data or its semantics. This is a machine-facing signal — provenance metadata, watermarking — described on the marking duty page.
- Article 50(4) — deployers, human-facing disclosure. A statement a person can perceive and understand. Machine-readable provenance in the file does not by itself discharge it, because nothing tells the viewer.
The practical implication is that a compliant workflow usually needs both: provenance that survives the pipeline, and a visible disclosure at the point of publication. The C2PA content credentials approach is the main candidate for the first and is discussed on the C2PA page; it is a candidate mechanism, not a legally mandated one, and the Act names no specific standard.
Timing, penalties and open questions
Article 50 applies from 2 August 2026 under Article 113, and that date survived the 2026 amendment described below. There is one date that is not 2 August 2026: for systems already placed on the market before that day, the marking obligations for AI-generated content apply from 2 December 2026, a four-month grace introduced by Regulation (EU) 2026/1744. It affects the Article 50(2) provider-side marking duty on an existing product, not the Article 50(4) deployer disclosure that is the subject of this page. Article 50(7) tasks the AI Office with encouraging and facilitating codes of practice at Union level to support effective implementation of the detection and labelling obligations, which is the mechanism through which the practical detail is expected to arrive. Non-compliance falls in the Article 99(4) tier: up to EUR 15 million or 3% of total worldwide annual turnover, whichever is higher.
Three things are genuinely open. What form of disclosure is adequate for content that travels — a label applied on one platform does not survive a re-upload elsewhere, and the Act places the duty on the deployer rather than on the intermediary. Whether the “does not substantially alter” exception in Article 50(2) covers common generative editing operations, which vendors read generously and the text does not resolve. And how Article 50(4) interacts with the Digital Services Act obligations on very large platforms, which run in parallel rather than in substitution — see the DSA page.
The AI Act was amended by Regulation (EU) 2026/1744, the digital omnibus on AI, published in the Official Journal on 24 July 2026 and in force from 27 July 2026. Two things it did are relevant here: it left the Article 50 transparency duties at 2 August 2026, and it gave systems already on the market before that date until 2 December 2026 for the marking obligations. It also postponed the high-risk regime (Annex III to 2 December 2027, Annex I products to 2 August 2028), which does not bear on Article 50(4). Anything else the amendment may have changed about Article 50 is outside what this page can confirm — read the consolidated AI Act text on EUR-Lex.
Top comments (0)