Updated 1 November 2026: rewritten to reflect that Article 50 now applies, the dates set by Regulation (EU) 2026/1744, and the final Code of Practice.
Article 50 of the EU AI Act requires providers of generative AI to mark outputs in a machine-readable, detectable format, and requires deployers to disclose deepfakes and certain AI-generated text on matters of public interest. It has applied since 2 August 2026. Generative systems already on the market before that date have until 2 December 2026 to meet the marking duty, under Regulation (EU) 2026/1744. A voluntary Code of Practice, published on 10 June 2026, has been assessed as adequate by the Commission and the AI Board.
Who has to do what under Article 50?
Article 50 divides the work between providers, who build systems, and deployers, who use them professionally.
- Article 50(1), providers of interactive AI: people must be told they are interacting with AI, unless that is obvious to a reasonably well-informed person.
- Article 50(2), providers of generative AI: synthetic audio, image, video and text must be marked in a machine-readable format and be detectable as artificially generated or manipulated. The technical solution must be effective, interoperable, robust and reliable as far as technically feasible.
- Article 50(3), deployers of emotion recognition or biometric categorisation: they must inform the people exposed to the system.
- Article 50(4), deployers: they must disclose deepfakes and AI-generated or manipulated text published to inform the public on matters of public interest.
- Article 50(5), all of the above: the information must be clear and distinguishable, given at the latest at first interaction or exposure, and accessible.
The site's risk tier guide explains how these duties sit alongside the Act's other tiers, and they can apply on top of high-risk duties (Article 50(6)).
Which dates apply after Regulation (EU) 2026/1744?
The Digital Omnibus on AI did not postpone Article 50, except for one transition for existing systems. Article 50 follows the Act's general application date, 2 August 2026. The Omnibus added Article 111(4): providers of systems generating synthetic content that were placed on the market before 2 August 2026 must comply with Article 50(2) by 2 December 2026.
That transition covers only provider marking. The Commission's Q&A confirms it and adds that content generated before 2 August 2026 does not need retroactive labelling, although the Commission encourages deployers to label it where possible. Deployers' duties under Article 50(3) and (4) therefore already apply.
Fines can reach EUR 15 million or 3% of worldwide annual turnover, with proportionality for smaller companies (Article 99(4) and the Q&A). National market surveillance authorities are the main enforcers. The AI Office's role is limited to cases such as systems built on a model from the same provider.
What counts as marking in practice?
The Act names no single technique, but the Code of Practice expects two layers for most media. Recital 133 mentions watermarks, metadata identifications, cryptographic methods, logging and fingerprints. Section 1 of the Code commits signatories to:
- digitally signed metadata, recording whether content is AI-generated or manipulated, for formats that can carry metadata; and
- an imperceptible watermark, except for very short text.
Plain text cannot carry metadata, so one layer is enough there. A single layer may also be enough for systems embedded in physical products with a technically controlled, closed environment. The AI Board said this carve-out may raise concerns about circumvention, but does not at this stage stop the Code from meeting its objectives. Providers also commit to offering detection mechanisms. A visible "made with AI" badge alone does not meet Article 50(2).
Article 50(2) has an exception for systems performing "an assistive function for standard editing" or not substantially altering the input. The Commission's guidelines (C(2026) 5054, non-binding) list examples such as standard colour and contrast adjustments. The Q&A adds that some outputs fall outside the duty, such as short sequences of numbers, symbols or letters, and describes a narrow exemption for business-to-business or industrial uses, under conditions set in the guidelines.
When must a business label content?
A business using AI must disclose deepfakes, in a lighter form for evidently artistic works, and must label AI-generated public-interest text unless a person has reviewed it and someone holds editorial responsibility.
A deepfake is AI-generated or manipulated image, audio or video that resembles existing persons, objects, places, entities or events and would falsely appear authentic (Article 3(60)). For evidently artistic, creative or satirical works, the duty is limited to disclosing the content's existence in a way that does not spoil the work.
Public-interest text covers published text informing the public on subjects such as politics, public health, justice or economic developments open to public debate, according to the Q&A. A marketing product description would not normally be caught, in this note's reading, unless it informs the public on such a matter.
Purely personal, non-professional use is outside Article 50. Under the guidelines, an activity that regularly brings an economic benefit is professional, and staff acting for a company fall under that company's duties. The Code offers an optional EU icon built around the letters "AI". The site's Article 50 kit has working templates.
What is the status of the Code of Practice?
The Code is final, voluntary and open for signature, and adherence does not by itself prove compliance. The timeline, from the Commission's page:
- 5 November 2025: kick-off plenary.
- 17 December 2025 and 3 March 2026: first and second drafts.
- 10 June 2026: final Code published.
- 8 July 2026: Commission Opinion C(2026) 4839 finds that the Code adequately covers Articles 50(2), (4) and (5).
- 9 July 2026: the AI Board's assessment reaches the same conclusion and says the Board will monitor implementation, in particular the interoperability of marking and detection.
- 31 July 2026: the Commission reports about 190 signatories. Its list, last updated on 24 September 2026, shows 95 for Section 1 and 192 for Section 2.
The Opinion allows providers and deployers to rely on the Code across the EU, "while recognising" that adherence is not conclusive evidence of compliance. It also says the Commission may assess the Code differently later. Those who do not sign must show compliance by other, equally adequate means, and the Q&A warns that they may face more requests for information. Article 50(7), as amended, still lets the Commission adopt common rules by implementing act if it finds a code inadequate.
What other law applies, and what about Türkiye and Korea?
Article 50 is only one layer, and its reach goes beyond the EU. A company outside the EU falls within the Act where a system's output is used in the Union (Article 2(1)(c)).
- EU bans: from 2 December 2026, systems generating non-consensual intimate imagery are banned under Article 5(1)(ba). See the note on the nudification ban.
- Personal data: a deepfake of a real person is likely to involve personal data, so the GDPR or the KVKK applies alongside.
- Türkiye: this note found no Turkish statute in force imposing a general duty to mark AI content comparable to Article 50. Companies selling into the EU still need to follow it.
- Korea: Article 31 of the AI Basic Act, in force since 22 January 2026, has its own notice and labelling duties. See the Korea guide and the note on further plans.
Frequently asked questions
When did Article 50 of the EU AI Act start to apply?
Article 50 of the EU AI Act has applied since 2 August 2026. Regulation (EU) 2026/1744 gives one transition: providers of generative systems placed on the market before that date have until 2 December 2026 to meet the Article 50(2) marking duty. Deployers' disclosure duties received no such extra time.
Is the Code of Practice on AI-generated content mandatory?
No. The Code of Practice on Transparency of AI-generated Content is voluntary. The Commission and the AI Board found it adequate on 8 and 9 July 2026, so signatories may rely on it to show compliance with Articles 50(2), (4) and (5), but adherence is not conclusive evidence. Non-signatories must show compliance by other adequate means.
Does AI-generated text always need a label under the EU AI Act?
No. Article 50(4) of the EU AI Act covers text published to inform the public on matters of public interest. It does not apply where the text has had human review or editorial control and a person holds editorial responsibility. Providers must still apply machine-readable marking under Article 50(2).
Is a visible "AI" label enough to comply with Article 50?
Not for providers. Article 50(2) of the EU AI Act requires machine-readable marking, and the Code of Practice expects signed metadata plus an imperceptible watermark for most media. A visible label is the usual way for deployers to meet their separate duty to disclose deepfakes under Article 50(4).
Sources: Regulation (EU) 2024/1689 and Regulation (EU) 2026/1744 as published in the Official Journal; Code of Practice on Transparency of AI-generated Content; Commission Opinion C(2026) 4839; the AI Board's conclusion on the Code; guidelines C(2026) 5054 and Commission pages, all read on 1 November 2026. The Board's text is undated; its date of 9 July 2026 comes from the Commission's page. General information, not legal advice.
