AI content marking: the 2 December 2026 deadline
Machine-readable marking of AI-generated content must be in place for pre-2026 generative systems by 2 December 2026. A practical pipeline check.
AI content marking has a real deadline for every generative system that predates August 2026: 2 December 2026. This guide explains what Article 50 asks of providers and deployers, why provenance quietly disappears in render, conversion and ingest steps, and how to keep both the machine-readable mark and the visible label intact in a template-driven pipeline.
What changes on 2 December 2026

Article 50 of the EU AI Act has applied since 2 August 2026. The 2 December date is not a second deadline for a new duty: it closes the only grace period the Commission describes. Providers of generative AI systems placed on the market before 2 August 2026 must comply with the machine-readable marking and detection duty in Article 50(2) from 2 December 2026. Everyone else is already in scope.
The short answer: by 2 December 2026, every generative system in your stack that predates August 2026 must write machine-readable provenance into its output — and your team stays responsible for the visible disclosure on any deepfake it publishes.
| Date | What applies | Who it hits |
|---|---|---|
| Before 2 August 2026 | Content generated earlier needs no retroactive label; the Commission encourages voluntary labelling | Deployers |
| 2 August 2026 | All Article 50 duties apply, including visible deepfake disclosure and AI-interaction notices | Providers and deployers |
| 2 December 2026 | Marking and detectability under 50(2) become mandatory for generative systems placed on the market before 2 August 2026 | Providers of those systems |
If you run automated image or video production, the calendar is the easy part. The hard question is where the mark physically lives — and that is usually the step your pipeline rewrites without asking.
Two duties that get confused
| Duty | Article | Who owes it | What it looks like |
|---|---|---|---|
| Machine-readable marking and detectability of synthetic output | 50(2) | Provider of the generative system | Provenance data or watermark in the file, readable by detection tools |
| Visible disclosure of a deepfake | 50(4) | Deployer, i.e. the team that publishes | Human-readable label at the latest on first exposure |
Article 50(2) does not apply "to the extent the AI systems perform an assistive function for standard editing or do not substantially alter the input data provided by the deployer or the semantics thereof". That is why a motion-graphics video built from your own layout data usually needs no AI label: the automation rearranges your content instead of inventing a person or an event.
The duties do not substitute for each other. The Commission's FAQ states that deployers cannot simply rely on the embedding a provider added to fulfil their own disclosure obligation. A viewer has to see the label without a provenance inspector.
A deepfake is AI-generated or manipulated image, audio or video content that resembles existing persons, objects, places, entities or events and would falsely appear to be authentic or truthful. For evidently artistic, creative, satirical or fictional works, the duty narrows to an appropriate disclosure that does not hamper the display of the work.
Where a machine-readable mark actually lives
C2PA Content Credentials are a signed manifest attached to an asset. The C2PA explainer describes assertions about origin, modifications and use of AI that are bound to the content by a cryptographic hash and signed by the software that produced it. A manifest is normally embedded in the file; it can also be linked through a soft binding such as an invisible watermark.
Article 50(2) asks for solutions that are "effective, interoperable, robust and reliable" as far as technically feasible, considering content types, implementation cost and the state of the art. That is why the Code of Practice on Transparency of AI-generated Content, finalised in June 2026 and assessed as adequate by the Commission and the AI Board, spends its first section on providers' marking rules rather than on label design.
The practical consequence for a content team: a mark is a property of a file, not of a brand. Your provider marks the file it hands you. The moment your pipeline produces a new file, that mark describes the input, not the output.
Why render pipelines are where marks disappear
Automated production is a chain of transformations, and provenance metadata is the lightest thing in the chain.
| Pipeline step | What happens to the mark | What to do |
|---|---|---|
| Generation in an AI tool | Manifest written by the provider | Keep the original; note tool and model version |
| Upload to DAM, CDN or CMS | Re-encoding or stripped ancillary data can drop the manifest | Inspect one asset after ingest |
| Template render | Output is computed and encoded again, so it is new bytes in a new file | Log the render as a derived asset; carry disclosure in the layout |
| PNG to WebP, MP4 to WebM | Container metadata may not survive | Test every format you ship |
| Thumbnail or social crop | No metadata at all | Do not treat it as a marked artifact |
None of this is malicious. A renderer's job is a correct frame, not someone else's signature. But "we generate with a tool that marks its output" is not an answer about the file you published last week. The render formats guide covers encoding trade-offs; provenance you have to check separately.
Put the visible label in the template, not in the export queue
Visible disclosure survives automation only if it is part of the layout contract. Written as typed inputs, the label becomes a field that either has a value or fails validation, instead of a manual step someone remembers at 17:20.
With typed templates, labelText can be a string input for localisation, labelIcon an image input for the EU icon set, and showLabel a boolean so a variant without synthetic content is excluded on purpose rather than left unlabelled by accident. Because a bulk batch is pinned to one immutable template version, placement also stops drifting between campaign waves — the same argument as in template input design.
| Constraint | Why it matters |
|---|---|
| Readability at final size | A 24 px label on a 1600 px master is unreadable in a feed crop |
| Contrast and placement | Keep it out of platform UI zones and safe areas |
| Duration on video | Show it from the first frame; the duty attaches at first exposure |
| Icon variant | The EU set ships as black, white, and both at 50% transparency |
The EU icons are optional; the labelling requirement is not, and an icon alone does not establish compliance. Section 2 of the Code of Practice holds the placement specifications for signatories.
Keep a provenance record you can hand over
You do not have to implement C2PA to be answerable for a campaign. You need three answers per asset: was it marked at intake, is the published file marked now, and which visible label did it carry?
| Field | Example value |
|---|---|
| Asset and render ID | render_01J… plus campaign key |
| Generator, model version, date | Tool, version, 2026-11-18 |
| Input reference | Prompt ID or template version |
| Mark present at intake | Yes / no / unknown, with inspection tool |
| Transformations | Ingest, resize, WebM conversion |
| Hash before and after | Two hashes, so an audit sees the change |
| Label variant | Icon plus text, bottom-left |
| Reviewer and date | Who signed off the creative |
Two hashes and one yes/no column turn a compliance argument into a diff. They also show you the day a provider changes how it marks output.
A one-hour audit you can run this week
- Pull ten AI-generated assets your team published in the last quarter and inspect each for a manifest.
- Push one original through your full pipeline — ingest, render, convert, publish — and inspect the result.
- Confirm every template that can carry synthetic content has a label block and a
showLabelinput. - Log findings with a date and tool versions. A dated record turns the deadline into a project.
What enforcement looks like
National market surveillance authorities enforce Article 50, with the AI Office covering systems built on general-purpose models where one entity provides both, or where the system sits inside a designated very large online platform. Fines can reach EUR 15 million or 3% of worldwide turnover for the preceding financial year, with proportionality for SMEs. This is an engineering and editorial guide, not legal advice.
FAQ
Does a C2PA manifest replace the visible deepfake label?
No. The manifest is the machine-readable layer owed by the provider; the visible disclosure is owed by the deployer, and the Commission states that the embedded mark alone does not satisfy it.
Do we have to relabel content created before 2 August 2026?
No retroactive obligation applies. The Commission encourages labelling older material where possible — usually an inventory task rather than a rendering task.
We only edit existing footage — are we in scope?
Partly. Assistive functions for standard editing, and systems that do not substantially alter the input or its semantics, fall outside the marking duty. If the result is a deepfake that could pass for authentic, disclosure applies to you as the deployer.
Is the deadline 2 August or 2 December 2026?
Both, for different parties. Article 50 applies since 2 August 2026; 2 December 2026 ends the grace period, for the marking duty of systems placed on the market before August 2026.
Sources
https://digital-strategy.ec.europa.eu/en/faqs/transparency-obligations-under-article-50-ai-act — European Commission FAQ on Article 50: application dates, grace period, deepfake disclosure, reliance on embedded marks — accessed 2026-11-20.
https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-50 — AI Act Service Desk, Article 50 consolidated text: standard-editing exemption, paragraphs 2, 4 and 5 — accessed 2026-11-20.
https://digital-strategy.ec.europa.eu/en/policies/code-practice-ai-generated-content — Code of Practice on Transparency of AI-generated Content: scope, sections, signatories — accessed 2026-11-20.
https://digital-strategy.ec.europa.eu/en/policies/eu-icons-labelling-ai-generated-content — EU icons for labelling AI-generated content: variants, use cases, optional status — accessed 2026-11-20.
https://spec.c2pa.org/specifications/specifications/2.2/explainer/Explainer.html — C2PA and Content Credentials explainer: manifest, assertions, content binding, soft binding — accessed 2026-11-20.
https://jsoncut.com/docs/templates/ — JsonCut typed templates and bulk rendering: input types, pinned template versions — accessed 2026-11-20.