Slop TVNewsLatest
Explainers

EU AI Act: legacy generative AI must mark output by December 2

The same date brings the EU's first bans on AI systems that generate non-consensual intimate material or child sexual abuse material.

Illustration: EU AI Act: legacy generative AI must mark output by December 2
Illustration: SLOP TV News. Logos are trademarks of their owners.

Key takeaways

  • Providers of generative AI systems placed on the EU market before August 2, 2026 must comply with Article 50(2) machine-readable marking by December 2, 2026, under Article 111(4) of the AI Act.
  • Article 5 prohibitions on AI systems that generate non-consensual intimate material or child sexual abuse material also begin to apply on December 2, 2026.
  • Article 50 is not new: the deployer duty to disclose deepfakes under Article 50(4) and the chatbot disclosure duty under Article 50(1) have applied since August 2, 2026.
  • Non-compliance with an Article 5 prohibition carries administrative fines of up to EUR 35,000,000 or 7% of worldwide annual turnover, whichever is higher.

On December 2, 2026, two separate obligations under the EU AI Act take effect. Providers of generative AI systems that were already on the EU market must have added machine-readable marking to what those systems output, and the EU's first prohibitions on AI systems that generate abusive sexual material begin to apply on the same day. Both dates come from the AI Act as amended by the Digital Omnibus on AI, Regulation (EU) 2026/1744, which the Official Journal published on July 24, 2026.

The AI Act is Regulation (EU) 2024/1689. Its transparency chapter, Article 50, has applied since August 2, 2026, and the Digital Omnibus did not move that date. It added a transition for one duty and it wrote two new bans into Article 5. Each piece lands on a different person, which is where most of the confusion sits.

What exactly happens on December 2, 2026

Two things, and they belong to different articles.

The first is a deadline. The Omnibus added a paragraph to Article 111 that reads: "Providers of AI systems, including general-purpose AI systems, generating synthetic audio, image, video or text content, that have been placed on the market before 2 August 2026 shall take the necessary steps in order to comply with Article 50(2) by 2 December 2026."

The second is a start date. The same regulation rewrote Article 113 so that Article 5(1), first subparagraph, points (ba) and (bb), together with the new Article 5(1a) and 5(1b), "shall apply from 2 December 2026", while the rest of Chapters I and II applied from 2 February 2025.

Neither item is a rewrite of the whole law. One is a four-month transition for the marking duty for systems already selling in Europe on August 2, 2026. The other switches on two bans.

Article 50(2): who has to mark, and with what

Article 50(2) is a provider duty, and it is separate from the deployer duties in the same article. It reads: "Providers of AI systems, including general-purpose AI systems, generating synthetic audio, image, video or text content, shall ensure that the outputs of the AI system are marked in a machine-readable format and detectable as artificially generated or manipulated."

The paragraph sets its own ceiling. It asks for technical solutions that are effective, interoperable, robust and reliable only as far as that is technically feasible, and it takes account of the limits of each content type, the cost of doing it, and the state of the art. It also exempts systems that only assist with standard editing or do not substantially change the input or its meaning.

Only legacy systems got extra time. A generative system first placed on the EU market on or after August 2, 2026 was expected to meet Article 50(2) from that day. The December 2 date in Article 111(4) covers the systems already on the market before August 2, 2026. Read plainly, the December deadline is a transition for older deployments, not the start of the marking rule.

The Commission's guidance on how to mark arrived in the summer. It published its Code of Practice on marking AI-generated content on June 10, 2026. The code is voluntary. IPTC's write-up of the code describes its first measure as two marking mechanisms used together: digitally signed and time-stamped metadata, and an imperceptible watermark embedded in the content. A second commitment asks signatories to make a detection system available so a member of the public can check whether a piece of content came from their system. Fingerprinting, or a registry lookup, is listed as an optional third mechanism.

The two new Article 5 bans

Article 5 already prohibited a list of practices, from social scoring to untargeted facial scraping. On December 2, 2026, it grows two more.

Point (ba) prohibits "the placing on the market, the putting into service or the use of an AI system that generates or manipulates realistic images, videos, audio or similar material of an identifiable natural person's intimate parts, or of an identifiable natural person engaged in sexually explicit activities, without that person's freely-given, specific, informed, unambiguous and explicit consent for that generation or manipulation".

Point (bb) prohibits "the placing on the market, the putting into service or the use of an AI system that generates or manipulates material or performance within the meaning of Article 2, points (c) and (e), of Directive 2011/93/EU, except where a 'without right' defence applies under national law". Directive 2011/93/EU is the EU's directive on combating the sexual abuse and sexual exploitation of children.

The Omnibus also wrote in two limiting paragraphs. Article 5(1a) says the provider-side ban bites only where the generation or manipulation is the system's intended purpose, or where the system's design, training, architecture, capabilities or user-facing features make it a reasonably foreseeable and reproducible outcome and the system lacks reasonable and adequate safeguards to prevent it. For deployers, the ban applies only where the deployer uses the system to generate that material. Article 5(1b) says a manipulation that does not increase the exposure of depicted intimate parts or change the nature of depicted sexual activity is not manipulation. The recitals add that the ban is not meant to stop providers from developing generative capabilities in general.

Provider duty and deployer duty are not the same thing

This is the distinction the December date makes easy to blur.

Article 50 holds duties for two roles. Providers must meet the marking duty in Article 50(2). Deployers carry the duties in paragraphs 1 and 4: the chatbot disclosure rule in Article 50(1), and the deepfake disclosure rule in Article 50(4), under which a deployer that generates or manipulates image, audio or video content constituting a deepfake must disclose that the content is artificially generated or manipulated and that disclosure is limited for evidently artistic, creative, satirical or fictional work.

Both of those deployer duties have applied since August 2, 2026, with no transition. The December 2 date in Article 111(4) reaches only Article 50(2), and only for systems placed on the market before August 2, 2026. A creator who publishes a deepfake in the EU is already inside the deployer duty today, and December 2 does not change that.

The penalty tier

A breach of an Article 5 prohibition sits in the top tier. Under Article 99(3), non-compliance with the prohibition of AI practices in Article 5 is subject to administrative fines of up to EUR 35,000,000 or, if the offender is an undertaking, up to 7% of its total worldwide annual turnover for the preceding financial year, whichever is higher. Transparency breaches under Article 50 fall in the next tier down, at up to EUR 15,000,000 or 3%.

What is still unknown

National supervision is not settled. The Commission has said the code of practice will be complemented by guidelines that clarify the scope of the obligations, and CASRAI, in its account of the December 2 obligations, notes it could not confirm how national authorities will enforce either provision or the status of harmonized standards on marking. The code is voluntary, so adherence is not the same as compliance.

The date should also not be confused with the high-risk deadlines, which move further out: Article 113 sets December 2, 2027 for high-risk systems under Article 6(2) and Annex III, and August 2, 2028 for those under Article 6(1) and Annex I.

What a creator should do before December 2

If you run your own generative model or ship one into the EU and it was live before August 2, 2026, the work now is to show how the system meets Article 50(2) by December 2. If you only use a vendor's tool, the marking duty is the vendor's, but you can ask where its detection mechanism lives and whether the outputs carry signed metadata, a watermark, or neither. If you publish a deepfake, the disclosure duty is on you, and it has been live since August 2, 2026.

The next dates to watch after December 2, 2026 are December 2, 2027 for the Annex III high-risk rules and August 2, 2028 for the Annex I rules.

The Commission's code of practice, its EU labeling icons and its Article 50 guidance are free to read at digital-strategy.ec.europa.eu.

Sources

  1. ai-act-service-desk.ec.europa.eu - Article 50 text, provider and deployer transparency duties
  2. ai-act-service-desk.ec.europa.eu - Article 111(4), the December 2, 2026 Article 50(2) date for legacy systems
  3. ai-act-service-desk.ec.europa.eu - Article 5(1) points (ba) and (bb) and paragraphs 1a and 1b
  4. ai-act-service-desk.ec.europa.eu - application dates for Chapters I and II and Article 5(1)(ba) and (bb)
  5. ai-act-service-desk.ec.europa.eu - penalty tier for Article 5 breaches
  6. eur-lex.europa.eu - Regulation (EU) 2026/1744, the Digital Omnibus on AI
  7. digital-strategy.ec.europa.eu - Commission press release, Code of Practice published June 10, 2026
  8. iptc.org - Code of Practice's two marking layers and public detection system
  9. casrai.org - confirmation that both obligations are fixed by Regulation (EU) 2026/1744