EU AI Act Article 50 deadlines for AI labels

The timeline

For retailers, the key date is 2 August 2026.

1. 12 July 2024: AI Act published in the Official Journal. Regulation (EU) 2024/1689 appears in the Official Journal of the EU.
2. 1 August 2024: AI Act enters into force. On the twentieth day after publication. Duties apply in stages; Article 50 only from 2 August 2026 (Article 113).
3. 17 December 2025: First draft of the Code of Practice. The Commission publishes the first draft of the code on transparency of AI-generated content.
4. 3 March 2026: Second draft of the Code. The Commission publishes the revised draft.
5. 8 May 2026: Draft guidelines. The Commission presents its draft guidelines on Article 50.
6. 9 June 2026: New York: GBL § 396-b applies. Advertising featuring AI-generated “synthetic performers” needs a conspicuous disclosure in New York where the advertiser has actual knowledge.
7. 10 June 2026: Code of Practice, final version. Voluntary code with the “AI” mark as main element, placement free of overlapping elements, and internal processes.
8. 8 July 2026: Commission finds the Code adequate. Opinion C(2026) 4839, published on 9 July. Following the Code does not create a presumption of conformity.
9. 20 July 2026: Guidelines C(2026) 5054. The Commission approves the content of its Article 50 guidelines. Formal adoption follows once all language versions exist.
10. 24 July 2026: Digital Omnibus published. Regulation (EU) 2026/1744 amends the AI Act, including Article 99 (small mid-caps) and Article 111 (transition for providers).
11. 27 July 2026: Omnibus enters into force. On the third day after publication. Articles 102 to 110 apply from this date, including the addition of the AI Act to the Representative Actions Directive (Article 110).
12. 29 July 2026: German KI-MIG enters into force. The German implementing act makes the Federal Network Agency (Bundesnetzagentur) the market surveillance authority and central complaints office.
13. 2 August 2026: Article 50 applies. Deployers must label deep fakes from this day, with no transition period (Article 50(4) and (5)). The Bundesnetzagentur accepts complaints through an online form.
14. 2 December 2026: Transition ends for providers. Generative AI systems placed on the market before 2 August 2026 must now mark their output in a machine-readable way (Article 50(2) with Article 111(4)).
15. 1 January 2027: California: duties for large platforms. Large online platforms must not strip provenance data and must make it available to users (Bus. & Prof. Code § 22757.3.1). In our reading, the Californian law does not impose duties on retailers.

Why retailers have no transition period

The Digital Omnibus (Regulation (EU) 2026/1744) amended the AI Act in several places in July 2026 but added only one transition rule for Article 50, which concerns providers of generative AI systems. The guidelines call it a “targeted grandfathering rule only with regard to the marking and detection obligations under Article 50(2)” (para. 153).

Deployers are subject to the general start date in Article 113. AI images in your store that are deep fakes have therefore needed a visible label since 2 August 2026. Whether an image is a deep fake is explained in deep fakes in online stores; the labelling check classifies individual images.

What 2 December 2026 means for retailers

Until 2 December 2026, AI tools that were already on the market before 2 August 2026 may deliver images without machine-readable marking (Article 111(4)). That has two consequences for you:

• Missing metadata does not prove that no AI was involved, not even after 2 December, because many editing steps lose metadata (see C2PA, IPTC and metadata).
• Your own duty stays the same. The visible label was required before, and you cannot rely on the provider's marking to meet it (para. 117).

Legacy images: what applies to older AI images

Deep fakes generated or manipulated before 2 August 2026 “do not need to be marked or labelled retroactively” (para. 154). The Commission nevertheless encourages deployers and anyone distributing such images to label them, without expecting disproportionate effort such as auditing entire content databases or changing product packaging that has already been printed.

Generation counts, not the upload

For images, para. 154 looks at when AI generated or changed them, not at the upload. An upload before 2 August 2026 does show that the image was generated before then. For later uploads you need other evidence, such as a timestamp in the metadata, a contractor's invoice or the project files.

New AI edit, new date

If you edit an older image with AI again after 1 August 2026, the date of the new edit counts, and the image falls under the duty if it is a deep fake.

Old image, new campaign: an open question

For text, para. 154 expressly looks at publication, so text generated before the deadline but published afterwards must be labelled. For images that sentence is missing, and on the wording, reusing an old deep fake in a new campaign does not create a new duty. No court or authority has confirmed this. We recommend a label whenever you put an old AI image to new use, which is also in line with the guidelines' wish to see legacy content labelled over time.

What comes next

• Formal adoption of the guidelines. According to the cover document it follows once all language versions exist; it had not been confirmed when we last checked. We will then re-check the paragraph numbers on this site (see guidelines).
• 2 December 2026. The transition ends for AI providers, and more images should then arrive with metadata.
• 1 January 2027. Duties for large platforms apply in California; for retailers in the EU this changes nothing about Article 50.

How the duty is enforced and which fines are possible is covered in fines and enforcement.