By Martin Scott

. 6 min read

On 2 December 2026, two things happen under the EU AI Act. Generative AI tools that were already on the EU market before 2 August 2026 lose their grace period and must mark their output as machine-made. On the same day, two new bans start: on AI systems that make non-consensual sexual images of real people and on systems that make child sexual abuse material.

That leaves about five weeks. The grace period is four months, shorter than the six the Commission first proposed, and it covers only one duty. Small software teams and agencies that ship their own branded generators are inside it.

What the law actually says

The marking duty sits in Article 50(2) of the AI Act. It covers providers of systems “generating synthetic audio, image, video or text content”. They “shall ensure that the outputs of the AI system are marked in a machine-readable format and detectable as artificially generated or manipulated.” Machine-readable means a mark that software can find, such as metadata (data stored inside the file) or an invisible watermark, not a visible label.

That duty has applied to new systems since 2 August 2026, according to the Commission’s AI Act implementation timeline. The Digital Omnibus on AI, the amending law passed this summer, added a transitional rule for older products. Article 111(4) applies to such systems “that have been placed on the market before 2 August 2026”. Their providers must comply with Article 50(2) “by 2 December 2026.”

The Commission’s original proposal, as described in the AI Act Service Desk FAQ, suggested “a transition period of 6 months” for this retrofit. The adopted text gives four months, from 2 August to 2 December. Any plan still built around a six-month window is working from the proposal, not the law.

What the grace period does not cover

The extension is narrow. The Commission’s guidelines on Article 50 say systems that are partly interactive and partly generative get the extra time “only with regard to the marking obligation under Article 50(2)”. The duty to tell people they are talking to an AI, under Article 50(1), has applied since 2 August 2026 with no extension.

There is one piece of relief. The same guidelines say outputs “generated or manipulated before 2 August 2026 do not need to be marked or labelled retroactively.” Old archives can stay as they are. New output from an old tool cannot.

Who counts as a provider

A provider is whoever develops an AI system, or has it developed, and offers it or puts it into use under its own name. The guidelines (PDF) give the example of a company offering “a chatbot, image generator, AI agent” under its own name. That company is the provider, “regardless of whether the AI system is provided for free or for payment” and wherever it is based.

This matters for small software teams and agencies that ship a branded image or copy tool built on someone else’s model. The guidelines say a provider “may rely on the marking solution implemented by an upstream model provider”. That reliance is “without prejudice to the responsibility of the provider of the AI system to demonstrate compliance.” In plain terms, if your vendor’s model already watermarks its output, you can use that. You still have to be able to show it works in your product.

Article 50(2) has its own exceptions. The duty does not apply where a system performs “an assistive function for standard editing” or does not substantially alter the input. The guidelines list grammar fixes, translation, format conversion, minor cropping and transcription as examples.

The guidelines also describe a narrow carve-out for internal business tools whose output is strictly technical, seen only by a defined group of staff, and never shared outside the company. All three conditions must be met.

What the code of practice asks for

The law does not name a technique. The Code of Practice on Transparency of AI-generated Content, published in final form on 10 June 2026, fills that gap. The Commission page says the Commission and the AI Board, part of the EU-level governance the Act sets up, have confirmed it “is an adequate voluntary tool” to show compliance.

The code text (PDF) commits signatories to “at least two layers of machine-readable marking” for audio, images, video and containerized text: digitally signed metadata plus an imperceptible watermark. Free-form text needs one layer. Free-form text of 200 tokens or fewer is exempt from watermarking. Tokens are the chunks of words a model reads and writes.

Signatories must also offer a free detection tool. Smaller players get a cost break. Those “with fewer than 1,000,000 monthly users” may charge a fee if detection is costly to run and one user sends heavy volumes.

Watermark detection must be interoperable by 2 February 2027. That February date is a commitment in the code, not a deadline in the regulation.

Take-up so far is modest. The Commission’s signatory page, last updated 24 September 2026, lists 95 signatories for the provider section and 192 for the deployer section. Deployers are the businesses that use the tools. It names Anthropic, Google, Meta, Microsoft, Mistral and OpenAI among provider signatories and says about half of all signatories are small and recent companies.

If you do not sign

The guidelines are blunt about the trade-off. Non-signatories should carry out “a gap analysis” against the code. They “may also be subject to a larger number of requests for information” from authorities. Adherence to the code can count as a mitigating factor when a fine is set.

The two new bans

The amended Article 5(1)(ba) bans AI systems that make “realistic images, videos, audio or similar material” showing an identifiable person’s “intimate parts” or that person “engaged in sexually explicit activities”. The exception is the person’s explicit consent. Point (bb) covers child sexual abuse material. Article 113 sets both to apply from 2 December 2026.

The ban is not limited to tools built for that purpose. Under Article 5(1a), a general system is also caught when such output is “a reasonably foreseeable and reproducible outcome”. That applies if it also lacks “reasonable and adequate technical safety measures” to prevent it. For anyone running an open image or video generator, filters and misuse reporting now carry legal weight.

What it costs to miss

Fines are set in Article 99. Breaching a prohibition can cost up to EUR 35 million or 7% of worldwide annual turnover, whichever is higher. Breaching the Article 50 transparency duties can cost up to EUR 15 million or 3%.

For small and medium-sized companies, Article 99(6) flips the rule: the cap is whichever of the two is lower. For a small studio, the realistic ceiling is the percentage of turnover. It is still an exposure that authorities can act on from 2 December.

Enforcement runs mainly through national market surveillance authorities, the regulators that police product rules in each country. The guidelines say the AI Office handles systems built on a general-purpose model from the same provider. We could not confirm whether any national authority has announced checks tied to the December date.

What to do this week

List every tool you offer under your own name that outputs images, audio, video or text. Note whether each was live before 2 August 2026. Then ask your model vendor in writing whether its output carries signed metadata and a watermark. Ask how to test that the marks survive your own processing, such as resizing or format conversion.

If a tool fails, pick one route now: add a third-party marking service, sign the code and follow its measures, or switch the feature off for EU users on 2 December. If you run an open image or video generator, document the safeguards that block sexual images of real people before the ban applies.

Sources

Martin Scott

Martin Scott is the editor for business and policy at AI Observer, covering startups and deals, regulation, health and science, and AI agency profiles. He follows the money and the legal text, and separates what a rule says from what people claim it says. His style is analytical and plain-spoken.

Spotted a mistake? Tell us on the corrections page.

Conversation

Every comment is reviewed before it appears, so there can be a short delay.