Almost all of Article 50 has applied since 2 August 2026. The one exception is Article 50(2), the duty on providers of generative AI to mark outputs in a machine-readable format: providers whose systems were already on the market before 2 August 2026 have until 2 December 2026. Chatbot disclosure, deep fake disclosure and the other duties did not move.
Key point: The 2 December 2026 date only covers Article 50(2) machine-readable marking, and only for providers of generative AI systems placed on the market before 2 August 2026. If you run a chatbot or publish deep fakes, your Article 50 duties already apply.
The dates, duty by duty
- Article 50(1): tell people they are interacting with an AI system. Applies to providers from 2 August 2026. This is the chatbot, voice agent and AI assistant rule.
- Article 50(2): mark synthetic audio, images, video and text in a machine-readable format. Applies to providers of generative AI systems from 2 August 2026. Systems placed on the market before that date have until 2 December 2026.
- Article 50(3): inform people exposed to emotion recognition or biometric categorisation. Applies to deployers from 2 August 2026.
- Article 50(4): disclose deep fakes, and AI-generated text published to inform the public on matters of public interest. Applies to deployers from 2 August 2026.
For what each paragraph requires in practice, see Article 50 of the EU AI Act, explained.
Where the 2 December date comes from
The AI Act (Regulation (EU) 2024/1689) originally applied Article 50 in full from 2 August 2026. The Digital Omnibus on AI, Regulation (EU) 2026/1744, changed that for one group. It was published in the Official Journal on 24 July 2026 and entered into force on 27 July 2026.
Article 111(4), inserted by the Digital Omnibus (Reg. 2026/1744), gives providers of AI systems that generate synthetic audio, image, video or text, including general-purpose AI systems, until 2 December 2026 to comply with Article 50(2), provided the system was placed on the market before 2 August 2026.
Three things follow from that wording:
- It is a provider rule. It helps the companies that build generative tools, such as image, video, voice and text generators. It does not give the businesses that use those tools extra time for anything.
- It is about marking, not disclosure. Article 50(2) is about machine-readable signals embedded in outputs, such as metadata or watermarks. Visible disclosure duties sit in Article 50(1) and 50(4), and those were not deferred.
- It only covers systems already on the market. A generative system placed on the market on or after 2 August 2026 had to comply with Article 50(2) from day one.
What the Digital Omnibus did not change
A lot of commentary summarises the Omnibus as "the AI Act was delayed". For most websites, that is the wrong takeaway.
- The big delays are for high-risk AI. Obligations for high-risk systems listed in Annex III moved to 2 December 2027, and for high-risk systems covered by Annex I product legislation to 2 August 2028.
- Article 50(1), 50(3) and 50(4) were not deferred. They have applied since 2 August 2026.
- The penalty regime applies. Breaches of Article 50 fall under Article 99(4): fines of up to €15,000,000 or 3% of total worldwide annual turnover, whichever is higher. For SMEs and start-ups the lower amount applies (Art. 99(6)), and Article 99(6a), added by the Digital Omnibus, extends that lower cap to small mid-caps. See what happens if you ignore Article 50.
Guidance that has landed since August
Two Commission documents now sit alongside the legal text. Neither is binding law, but both show how Article 50 is expected to be applied.
- Commission guidelines on Article 50, adopted on 20 July 2026. They set out the Commission's reading of each paragraph. For example, law firms report that the guidelines expect interaction disclosure at the very first interaction and reject disclosures hidden in terms and conditions.
- The Code of Practice on marking and labelling AI-generated content, published in its final version in June 2026. It is voluntary. It describes layered machine-readable marking for providers (for example, metadata combined with watermarking) and visible labelling for deployers, including a common EU icon.
Following the Code of Practice is not the same as complying with Article 50, but it is likely to become a practical reference point.
What to do before 2 December
If you only use AI tools (a chat widget, an image generator, an AI writing tool):
- Don't treat 2 December as your deadline. Check your chatbot disclosure and any deep fake or public-interest text disclosures now.
- Ask the vendors of your generative tools whether their outputs carry machine-readable marking, and whether your workflow (resizing, re-exporting, compressing) strips it.
- Label deep fakes you publish, such as realistic AI images, audio or video of real people, places or events. See deepfake disclosure.
If you build or sell a generative AI system:
- Work out whether you are the provider of a system that generates synthetic content, and when it was first placed on the market.
- If it was on the market before 2 August 2026, plan machine-readable marking of its outputs by 2 December 2026. If it launched later, the duty already applies.
- Use the Code of Practice as a practical reference for marking techniques, alongside the legal text.
For the wider picture, see AI-generated content labelling under Article 50.
A free Article50.io scan checks one public page of your site for two things: a detectable AI chat widget with no AI-disclosure wording on the page or in the widget as it loads (flagged under Article 50(1)), and machine-readable AI-generation markers in the page markup with no visible disclosure (flagged under Article 50(4)). It does not test whether a generative tool's outputs carry Article 50(2) marking, and a clean scan does not mean you are compliant.
Frequently asked questions
Did the Digital Omnibus delay Article 50?
Only one part of it. Article 111(4), inserted by the Digital Omnibus on AI (Reg. 2026/1744), gives providers of generative AI systems already on the market before 2 August 2026 until 2 December 2026 to comply with Article 50(2) machine-readable marking. Chatbot disclosure (Art. 50(1)), emotion recognition and biometric categorisation notices (Art. 50(3)) and deep fake and public-interest text disclosure (Art. 50(4)) have applied since 2 August 2026.
Does the 2 December 2026 deadline apply to my website chatbot?
No. Telling people they are talking to an AI is Article 50(1), which has applied since 2 August 2026. The 2 December 2026 date only covers Article 50(2) machine-readable marking by providers of generative AI systems that were already on the market before 2 August 2026.
I publish images made with an AI image generator. Is 2 December my deadline?
Not for anything you do yourself. Machine-readable marking under Article 50(2) is the image generator's provider duty, and the 2 December date is a provider transition. If an image you publish is a deep fake, meaning realistic content that resembles real people, places or events and would falsely appear authentic, your disclosure duty as a deployer under Article 50(4) already applies.
Are the Commission's Article 50 guidelines binding?
No. The guidelines adopted on 20 July 2026 set out the Commission's interpretation, and the Code of Practice on marking and labelling AI-generated content is voluntary. Only the AI Act itself is binding, and only the Court of Justice of the EU can give a definitive interpretation. In practice, regulators are likely to look at both documents.
What are the fines for breaching Article 50?
Under Article 99(4), up to €15,000,000 or 3% of total worldwide annual turnover, whichever is higher. For SMEs and start-ups the lower of the two applies (Art. 99(6)), and Article 99(6a) extends that lower cap to small mid-caps.
This article is general information, not legal advice.