What has to be done by December 2, 2026?

Machine-readable marking on AI-generated images, audio and video produced by generative systems that were already on the market before August 2, 2026. That is the end of the EU AI Act Article 50 grace period for legacy systems. Human-perceptible disclosure — the visible label — was already required from August 2, 2026 with no grace period, and systems launched after that date got no grace period on either limb.

What is the minimum viable compliance stack?

Four pieces: a visible or audible disclosure rendered at output time, a C2PA Content Credentials manifest signed at generation, a durable embedded watermark underneath the manifest because manifests are stripped by re-encoding, and — if California's SB 942 applies to you — a free public detection page anyone can use to check a file. Add a published training-data summary if you are a model developer rather than only a deployer.

How long does implementation take?

Manifest signing is typically one to three engineering weeks if your media pipeline has a single choke point, and considerably longer if generation is spread across services. The free public detection tool required by SB 942 is the long pole because it is a hosted product with its own uptime and abuse-handling requirements, not a library call. Start with the pipeline audit — you cannot scope any of it until you know every path that emits generated media.

What happens if a user strips the watermark?

Both regimes place the obligation on the provider to mark at generation, not to guarantee the mark survives arbitrary downstream tampering. Marking robustly and documenting your method is the defensible position. This is also why layering matters: a C2PA manifest alone disappears on re-encode, so pairing it with a durable pixel or audio watermark is what makes a technically-feasible claim credible.

Quick Answer

How to Meet the Dec 2, 2026 AI Content Marking Rule

Published: September 6, 2026

The Short Answer

Deadline: December 2, 2026. That is when the EU AI Act Article 50 grace period ends for machine-readable marking on generative systems that were already on the market before August 2, 2026.

Everything else in this area is already live: the visible-disclosure limb of Article 50 and California’s SB 942 both took effect August 2, 2026, and California’s training-data transparency law AB 2013 has been in force since January 1, 2026.

So if you are reading this in September 2026, you are not preparing for a future deadline. You are closing one gap while two others may already be open.

Step 1 — Audit Every Path That Emits Generated Media

You cannot scope this until you know the surface area. Enumerate:

Every endpoint that returns a generated or substantially altered image, audio file or video
Every batch or async job that writes generated media to storage
Every third-party model you call whose output you redistribute under your own brand
Every editing feature that materially alters real media (inpainting, voice cloning, face swap, background replacement) — “substantially altered” catches these

Completion criterion: a written list of every emission path with an owning team and a yes/no on whether it currently marks output. Most teams find two to four paths they had forgotten.

Step 2 — Ship the Visible Disclosure First

This limb had no grace period. If it is missing, you are already non-compliant, and it is also the cheapest thing to fix.

Images: an on-canvas badge or corner label, burned in at render time, not a CSS overlay that vanishes on download
Video: a persistent or opening-frame indicator
Audio: a spoken or tonal notice, or a clear disclosure in the delivering interface

⚠️ The common failure: labelling in the UI rather than in the artifact. A caption under an image in your web app does not travel with the file. Article 50 concerns the content, not your product chrome.

Completion criterion: download a generated file from every path in Step 1 and confirm the disclosure is present in the file itself.

Step 3 — Sign a C2PA Manifest at Generation

This is the machine-readable limb, and the piece with the December 2 date on it.

Do it at the choke point. If every generation route funnels through one media service, you add signing once. If it does not, consolidating first is usually cheaper than instrumenting five pipelines.

What goes in the manifest: the generating system, the timestamp, the nature of the operation (generated vs. altered), and a signature chaining to a certificate you control.

What to decide up front:

Decision	Guidance
Certificate issuer	Use a recognised issuer; self-signed manifests verify but carry no trust weight
Key custody	HSM or managed KMS — a leaked signing key invalidates every manifest you ever issued
Rotation	Plan it now; retrofitting rotation into a signing service is painful
Retention	Keep issuance logs — they are your evidence of compliance

Completion criterion: a validator reads a correct, chain-verifying manifest from output produced by every path in Step 1.

Step 4 — Layer a Durable Watermark Underneath

⚠️ A C2PA manifest is metadata and metadata dies. Screenshot the image, transcode the video, strip EXIF — the manifest is gone. Worse, absence proves nothing: “never marked” and “mark removed” are indistinguishable to a verifier.

That is why “where technically feasible” in Article 50 reads poorly if you shipped only the strippable layer. Pair the manifest with an embedded watermark that survives common transformations — Google’s SynthID is the widely cited example; several vendors ship equivalents, and there are open implementations for image and audio.

Layer	Survives re-encode?	Survives screenshot?
C2PA manifest	❌	❌
Embedded watermark	✅	usually ✅
Visible label	✅	✅ unless cropped

Completion criterion: re-encode a marked file at reduced quality and confirm your detector still finds the watermark.

Step 5 — Build the Free Public Detection Tool (California)

If SB 942 / AB 853 applies to you, this is a hard requirement and the longest lead time in the whole programme, because it is a product rather than a library call:

Free. No paywall, no enterprise-sales gate.
Publicly accessible. Anyone, not just your customers.
Accepts an upload and reports whether your system generated or substantially altered it.

Which means you also own: hosting, rate limiting, abuse handling, an upload path that will be pointed at hostile files, and a privacy posture for content strangers hand you. Do not treat uploaded files as trusted, and do not retain them longer than you need.

Completion criterion: a logged-out browser on a clean network can upload a file and get a verdict.

Step 6 — Publish the Training-Data Summary (Developers Only)

If you develop a generative model or service rather than only deploying someone else’s, California AB 2013 has required a public, website-hosted summary since January 1, 2026: dataset sources and owners, number and types of data points, and whether the data includes copyrighted or personal information.

New York’s A 6578 runs a parallel track, and the federal TRAIN Act, introduced January 2026, would add copyright-holder inspection rights — treat that one as pending, not operative.

Completion criterion: a public URL, linked from your product or trust pages, that a regulator could find without asking you.

Step 7 — Write Down What You Did

Compliance you cannot evidence is compliance you cannot defend. Keep, in one place:

The Step 1 emission-path inventory, dated
Which marking layers each path applies
Your robustness testing results from Step 4
Certificate and key-management procedure
The date each control went live

Completion criterion: a single document a reviewer can read in ten minutes.

Sequencing If You Are Starting Now

With roughly twelve weeks to December 2:

Weeks	Work
1	Step 1 audit; Step 2 visible labels shipped (already overdue)
2–3	Certificate issuance, key custody, signing service design
4–6	Step 3 manifest signing live on all paths
5–8	Step 5 detection tool — start early, it is the long pole
7–9	Step 4 watermark layer and robustness testing
10–12	Step 6 disclosures, Step 7 documentation, buffer

The two things that go wrong: teams start with the interesting cryptography and discover the detection tool in week nine, and teams scope against one pipeline and find three more in testing. Do the audit first.

Last verified: September 6, 2026.