The December 2026 AI Act marking transition: who it affects

2 December 2026: the marking transition ends

Providers of generative AI systems that were on the market before 2 August 2026 have until 2 December 2026 to implement machine-readable marking of outputs (Article 50(2)). After that date, the distinction between older and newer systems disappears for this duty.

This is narrower than 'the AI Act starts in December'. Everything else in Article 50, and all of Article 4, has been enforceable since August.

Who does not get a December extension

A business that only uses someone else's generative AI system does not inherit the provider's machine-readable marking duty. Its own Article 50 duties depend on what it deploys or publishes: chatbot interaction, emotion-recognition or biometric-categorisation exposure, deepfakes, and specified public-interest text.

The transition does not postpone Article 4 literacy measures, chatbot disclosure, deepfake disclosure or the other Article 50 duties that already apply. December is not a second general start date for the Act.

The winter to-do list

If you provide a generative AI product that was already on the market before 2 August 2026, finish machine-readable marking before 2 December. Every other business should classify its role and content instead of treating that provider transition as its own deadline.