The AI Act marking grace period and its limits
One Article 50 duty got a runway to 2 December 2026. Here is exactly which duties had no runway at all and were binding from day one.
The 2 December 2026 date has been quoted in enough summaries to have acquired a meaning it does not have: that Article 50 was postponed, that nothing bites until winter, that there is a quiet four months before anyone has to change anything. One obligation got a runway. It is the obligation that sits furthest from most people reading about it, and it is the only one on the list.
This is a duty-by-duty ledger of what was live on 2 August 2026 with no runway at all — written as a planning document, not as legal advice.
The ledger
| Duty | Binds | Applied from | Runway |
|---|---|---|---|
| Machine-readable marking of synthetic output (Art. 50(2)) | Providers of generative AI systems | 2 Aug 2026 | Until 2 Dec 2026 for systems already on the market |
| Perceivable labelling of deepfakes at first exposure (Art. 50(4)) | Deployers publishing the content | 2 Aug 2026 | None |
| GPAI copyright policy and training-content summary | Providers of general-purpose AI models | 2 Aug 2025 | Commission enforcement powers from 2 Aug 2026 |
| Code of Practice on Transparency of AI-Generated Content | Signatories only | Finalised 10 Jun 2026 | Voluntary throughout |
| Latent provenance marking, California AI Transparency Act | Covered GenAI providers above one million monthly users | 2 Aug 2026 | None |
Read the runway column top to bottom. Exactly one cell contains a date in the future, and the row it sits on is a duty on model providers. Every duty that lands on the person publishing content was binding from the first day.
The 2 December backstop itself traces to the Omnibus package agreed by the EU co-legislators and set out in the Commission's own FAQ material, targeted at the marking and detection obligations for generative systems placed on the market or put into service before 2 August 2026. Treat it as the working date to plan against and confirm final adoption status with counsel if a campaign carries meaningful EU exposure. What it is not is a general deferral of Article 50.
What the runway attaches to
The grandfathering attaches to a system, not to a company, an account, a project or a file. The question it answers is whether a particular generative AI system was placed on the market or put into service before 2 August 2026 — in which case its provider has until 2 December to bring the marking and detection machinery into conformity.
Three consequences follow that are easy to get backwards.
A grandfathered system does not grandfather your publication. If you publish a realistic synthetic clip made on a tool that is inside its marking window, the deployer disclosure duty on you is unaffected. The tool's runway is the tool's runway.
A newer system in the same product may not be inside the window. Providers ship new models continuously. "This platform existed in July" and "this specific system was on the market in July" are different statements, and the second is the one the rule turns on.
Your back catalogue is not covered by it either. The runway is about bringing a system into conformity going forward. It is not a retrospective amnesty for content, and it says nothing about material you are still actively distributing.
Why the marking duty got a runway and the labelling duty didn't
The asymmetry makes sense once you look at what each obligation actually requires someone to build.
The deployer label is a publishing decision. Deciding to put a readable line on a video, and putting it there, is work measured in minutes. There is nothing to engineer.
Machine-readable marking is infrastructure. It means embedded provenance metadata or a technical signal that survives export, detectable by tooling built to look for it — across every output type a platform supports, retrofitted into a product that was not designed around it. It also has to survive the real world, which is the part that makes it hard: a manifest is bound to a file, and re-encoding, format conversion, screenshotting or editing in a non-aware tool destroys or invalidates it. Durable approaches stack an invisible watermark and a perceptual fingerprint alongside the manifest to survive recompression, and even that combination is defeated by heavy crop or deliberate removal. Where SynthID-style marking stops working is the honest version of that ceiling.
Retrofitting all of that across an existing catalogue is a four-month engineering programme, not a checkbox. Which is precisely why the runway exists, and precisely why it was not extended to a duty that requires no engineering at all.
The three misreadings that cost something
"Nothing bites until December." The deployer labelling duty has been binding since 2 August 2026 with no grandfathering attached. So has California's provider-side marking duty, which arrived the same day and had no equivalent runway — two transparency laws, one start date covers how the two regimes interact for anyone selling into both markets.
"My tool isn't watermarking, so this clearly doesn't apply yet." The absence of a visible or embedded marker tells you where that provider is on its own conformity timeline. It tells you nothing about your obligations. Reading a vendor's engineering schedule as a compliance signal is the single most expensive mistake available here, and it is entirely avoidable.
"Once marking switches on, my disclosure job is done." Different artefacts, different readers. Embedded marking answers a detector. Your label answers a viewer at first exposure. Neither absorbs the other — the split is laid out in provider or deployer: who labels under Article 50.
What to actually do between now and December
The practical risk in this window is not enforcement. It is that the files you are shipping right now will contain different things in November than they did in August, and you will have no record of which was which.
- Inventory the generation tools in your chain. For each one, note whether the specific system you use was on the market before 2 August 2026. That is the set whose export behaviour is most likely to change without an announcement.
- Re-check what your exports contain, monthly. Marking behaviour can switch on mid-campaign. Once California's covered providers stand up free public detection tools under their own obligation, checking becomes something you can do rather than something you assume — auditing your exports with detection tools covers that workflow.
- Keep disclosing regardless. Your synthetic media disclosure has been due since 2 August. Nothing in this window changes it.
- Keep a dated record of what you generated. A request like "save this week's avatar clips and voiceovers to a project called 'EU campaign — August 2026'" creates a timestamped trail through saving generations to a project. It is a production habit, not a compliance product, and it makes a real compliance conversation much shorter.
- Diary 2 December. Not because your duties change that day, but because your suppliers' do, and the files you receive after it should look different.
- Rebuild the export step once, not per project. If your finishing chain strips metadata — most re-encodes do — that is a pipeline defect worth fixing while you have a reason to look at it. Rebuilding your export step is the practical version.
FAQ
Does the grace period mean I can stop disclosing until December?
No. The runway applies to the provider-side marking and detection obligation under Article 50(2). The deployer-side labelling duty under 50(4) has applied since 2 August 2026 with no equivalent extension.
How do I know if a tool I use is inside the window?
Ask the provider whether the specific generative system was placed on the market or put into service before 2 August 2026, and what its marking roadmap is. It is a fair procurement question and increasingly a normal one.
Is 2 December 2026 definitely the date?
It is the date set out in the Commission's FAQ material describing the targeted grandfathering agreed in the Omnibus package, contingent on adoption as agreed. Plan against it; verify final status with counsel if you have real exposure.
Did anything else in the AI Act have a 2 August 2026 date?
Yes — the Commission's enforcement powers over the general-purpose AI model obligations start then, though those substantive duties, including the copyright policy and training-content summary, have applied since 2 August 2025. That is a separate track from Article 50 and lands on model providers rather than on content producers.