← Back to blog
AiAbout 7 min read

December 2 Is Not a Postponement: the EU Marking Deadline Reaches Models Already on Sale

Published Nov 14, 2026
December 2 Is Not a Postponement: the EU Marking Deadline Reaches Models Already on Sale

There is a date in the EU AI Act calendar that most published timelines skip. It sits between the transparency obligations of August 2026 and the high-risk rules of December 2027, and for a large number of companies shipping image and video generation, it is the only date that matters this quarter.

On December 2, 2026, the transition period for marking AI-generated content closes.

The provision is Article 50(2) of Regulation (EU) 2024/1689. Providers of systems that generate synthetic audio, image, video or text must ensure the output is marked in a machine-readable format and is detectable as artificially generated or manipulated. The technique is not prescribed. The Act asks for marking that stays effective across tools and survives real pipelines as far as technically feasible, which reads like a standard written by people who knew no single method survives every distribution path.

Article 50 as a whole has applied since August 2, 2026. One duty inside it got extra time: the marking requirement, and only for generative systems that were on the market before that date. Article 111(4) sets the end of that grace at December 2. The European Commission's own guidance is blunt about the scope. Those providers must comply only as from December 2, 2026, which means anything launched after August 2 should have been marking from day one.

The misreading that is going around

The AI Omnibus, Regulation (EU) 2026/1744, in force since July 27, moved the high-risk deadlines. Stand-alone Annex III systems slipped to December 2027. High-risk AI embedded in regulated products slipped to August 2028.

That revision created a general impression that Europe had softened. It has not, at least not here. The Omnibus left the prohibitions, the general-purpose model rules, and the Article 50 transparency duties exactly where they were. The December 2 date arrives with the Omnibus in force, not in spite of it.

There is a second, sharper misreading. A team that shipped a generative feature in 2024 or 2025 and read "December 2026" in August may have concluded it had a few months of runway. It had a few months of runway on one clause. Every other part of Article 50 has been binding since August, including the duty to tell people they are talking to an AI system, and including the deployer duties that carry no transition at all.

Four duties, two columns

Article 50 splits by role, and the split is where most compliance plans go wrong.

If you build the system, you owe two things. Under 50(1), users must be told they are interacting with an AI system unless that is obvious. Under 50(2), synthetic output must carry a machine-readable mark.

If you use the system under your own authority, you owe three. Under 50(3), you disclose emotion recognition and biometric categorisation. Under 50(4), you disclose deepfakes and AI-generated text published on matters of public interest. Under 50(5), the disclosure happens at the first interaction, and it has to be clear and distinguishable.

The same piece of content can trigger both columns. A marketing team that generates a synthetic presenter, dubs them, and publishes the result is a provider in one part of the workflow and a deployer in another. Companies that bought a single labelling toggle and considered the matter closed have usually satisfied neither column properly.

One technical point deserves emphasis because it is widely assumed to be solved by marking. A machine-readable mark does not discharge the deepfake duty. Article 50(2) is aimed at machines that will read the metadata. Article 50(4) is aimed at a person who will watch the video. A C2PA manifest on a file nobody inspects does not tell an audience that the speaker is not real.

Why the hard part is scoping

The engineering, for most teams, is a week of work with an existing library. The scoping is where December gets tight.

Start with an inventory question that sounds trivial and is not: which of your systems were on the EU market before August 2, 2026, and which are new? The answer determines the deadline, and for products with continuous deployment the answer is genuinely ambiguous. A model released in 2025 and updated weekly with new weights is arguably a new system every week. A feature flag that turns on generation inside an existing product is arguably a new system on the day it turns on.

Then decide which outputs are in scope. The Commission's guidance carves out more than most summaries report. Source code is out. Machine-to-machine output is out. Short strings are out. Closed-loop industrial output is out. Standard assistive editing is out, and translation is named as an example.

That last exclusion is doing a lot of work in a market where half the products are editors. A tool that removes a background or adjusts exposure is assisting; a tool that invents a person who was not in the frame is generating. The line is whether the system substantially alters the input data or its semantics, and it will be argued case by case, which is exactly the kind of line that a compliance officer should write down before a regulator asks.

Layering, and the reason for it

No single marking technique has been recognised as sufficient across audio, image and video. The guidance points toward a layered approach, and the reason is mechanical.

Provenance metadata travels with a file and survives most professional pipelines. It does not survive a screenshot, a re-encode, or a platform that strips EXIF on upload, which describes most social platforms. Watermarking survives the re-encode and dies under cropping, heavy compression, or a phone pointed at a screen. Invisible watermarks embedded in the latent space survive more than either, and cost more to implement and to verify.

The layering answers a mechanical fact: distribution destroys evidence, and the platform that receives your file is the least controllable part of your pipeline.

A small gold foil seal pressed onto the edge of a sheet of glass.

What the December date costs if you miss it

Transparency breaches sit in the lower penalty tier: up to 15 million euros or 3 percent of global annual turnover, whichever is higher. The two new prohibitions that also start on December 2 sit in the top tier at 35 million euros or 7 percent, and they cover AI systems that generate child sexual abuse material and non-consensual intimate imagery.

The prohibitions are the reason this date is not a paperwork exercise. Any image or video pipeline that accepts arbitrary user input has been one prompt away from producing prohibited material since it launched. Until December 2, that exposure was governed by general law in most member states. After December 2, it is a named prohibited practice in the Act, in the highest penalty band, and the provider is the party on the hook.

A realistic three-week plan

Write the system inventory. Product name, launch date, whether generation is on by default, which formats it outputs. Two pages.

Map each output to a duty, and mark the exclusions you are relying on with a one-line justification. A memo that says "this function only assists standard editing and does not alter semantics" is worth more in an inquiry than a policy document that says the company takes transparency seriously.

Pick your marking stack and test it against your own distribution path rather than against a spec. Run your output through the platform where it will actually be posted, then check whether the provenance survived. Most teams discover at this step that their pipeline's last mile removes the evidence their first mile added.

Then tell people. The disclosure duty at first interaction has been live since August, and it is the cheapest item on the list.

None of this is glamorous, and none of it is optional after the second of December. The companies that will find the date easy are the ones that treated marking as a product feature rather than a compliance report, because a visible, explainable label is also the thing that makes a synthetic presenter usable in a market that does not entirely trust synthetic presenters.

Related articles

Google replaced the model behind Search, and the number that moved was a shell benchmark

Gemini 4 Hydrogen became the default on 12 November, taking over from Gemini 3.8 Flash across Search, the Gemini app and the developer API. It runs about 2.5 times faster, scores 57% on Terminal-Bench 4.1 where its predecessor scored 18.2%, and costs $1.50 in and $7.50 out per million tokens.

Three ways a coding agent wastes money, measured on 30 repositories

TokenEconomizer watches what an agent did in a repository you work in often, then stops it from rediscovering the same build command, loading tools it never uses and looping on one error. Across five agent and model combinations it cut mean cost by 18.4% to 30.9%, cheaper in all 30 repository cells, with Pass@1 up in four of five.

earshot gives coding agents 922 sounds, and every one is rendered from code

An agent can write a game, an interface or a film and never hear any of it. earshot answers requests like "three variations of a footstep on gravel" from a catalogue of 922 designs built from public-domain recordings, with no audio files in the repository at all. It runs on CPU, as a CLI, a Python package, an MCP server and a Claude Code plugin.

The agents stealing API keys were not sophisticated. The gateways were open.

Amazon's decoy servers spent October posing as inference APIs and logged agents built on ARTEX pulling provider keys out of LiteLLM, Dify, Langflow and Open WebUI. The way in was a default master key and a half-finished install. ARTEX itself ships with budgets, scope gates and evidence for every finding.