EU AI Act grace period ends 2 December for AI disclosure and content marking

The 2 December cutoff

Article 50 of the EU AI Act, the provision covering transparency towards users, has applied since 2 August 2026 to any product that interacts with people in the EU, according to a post on dev.to. Systems that were already on the market when that date passed received a grace period for one specific duty — machine-readable marking of AI-generated output — and that grace period ends on 2 December 2026. From that point, the marking obligation applies to established products as well as new ones.

Three separate obligations

According to the dev.to post, Article 50 bundles three distinct requirements:

• Interactive systems such as chatbots must disclose that they are AI, either before the first interaction or during it.
• AI-generated text, images, audio and video must carry a mark that machines can read, for example embedded metadata or a watermark.
• Deepfakes additionally require a disclosure label that is visible to the user, not merely machine-readable.

The first and third are interface decisions: what the disclosure says and when the user sees it. The second is an engineering task, and it is the one covered by the December deadline for pre-existing systems.

Who the rules reach

The scope is broad, the post argues: any chat feature, AI content tool, onboarding bot, support agent, voice interface or avatar that an EU user might encounter. The obligations follow the users rather than the company's headquarters, so a team with no EU office or legal entity is still caught if people in the bloc can use the product. The penalty ceiling the post cites is 3% of global turnover.

A practical checklist

The post reduces the work to four steps:

1. Map every AI feature that interacts with users, including chat, content generation, voice and avatars.
2. Place disclosure copy at the right moment — at or before the first interaction for chatbots, visible labels for deepfakes.
3. Mark generated content in machine-readable form. Because the technical standard is still being finalised, the post recommends implementing an interim approach, such as metadata or watermarking, and documenting it.
4. Keep records of what is disclosed and where, since regulators ask for evidence rather than intentions.

The unfinished standard is the awkward part. Teams have to build marking pipelines now, without a final specification to target, which makes the paper trail — documenting the interim method and why it was chosen — as important as the mark itself. It is worth noting that the dev.to post also promotes a paid template pack from its author; the obligations and dates it describes, however, come from the Act itself.

Why it matters

Most AI Act discussion centres on model providers and large platforms, but Article 50 lands squarely on ordinary product teams. If you ship a support bot, an AI writing feature or generated media to users anywhere in the EU, disclosure and marking are now baseline product requirements, not legal afterthoughts. The 2 December date removes the last transition cover for systems built before August, and the extraterritorial reach means having no EU office offers no shelter. With fines expressed as a percentage of global turnover, and with regulators asking for evidence rather than intentions, the cheapest compliance is the work done and documented before the deadline — not the scramble after a first enquiry.