What is the difference between EU AI Act Article 50, California SB 942 and C2PA?

Article 50 and SB 942 are legal obligations; C2PA is the technical standard most providers use to satisfy them. EU AI Act Article 50 became applicable August 2, 2026 and requires both a human-visible disclosure on AI-generated or manipulated images, audio and video and a machine-readable marking where technically feasible. California's SB 942, extended by AB 853, took force the same day and requires covered providers to publish a free public detection tool plus embed a latent, machine-readable provenance record. C2PA Content Credentials is the cryptographically signed manifest format that carries that provenance.

What is the December 2, 2026 AI watermarking deadline?

It is the end of the EU AI Act Article 50 grace period for machine-readable marking on generative systems already placed on the market before August 2, 2026. Visible labels were required from August 2, 2026 with no grace period; the embedded machine-readable marking on pre-existing systems must be in place by December 2, 2026. Systems launched after August 2, 2026 had no grace period at all.

Does C2PA compliance satisfy the EU AI Act?

C2PA Content Credentials can satisfy the machine-readable marking limb of Article 50, but not the whole article. Article 50 also requires a human-perceptible disclosure — a visible label or audible notice a person can notice without tooling — and C2PA manifests are invisible by design. You need both. C2PA manifests are also strippable by re-encoding, so most providers pair them with a durable watermark such as SynthID.

Do these rules apply to a US company with no EU office?

Article 50 applies on a market basis, not an establishment basis — if your generative output reaches users in the EU, the obligation attaches regardless of where you are incorporated. California's SB 942 applies to covered providers offering generative systems to California residents above its user threshold. In practice most companies of any scale end up in scope of both and build to the stricter combined requirement.

Quick Answer

EU AI Act Art. 50 vs California SB 942 vs C2PA 2026

Published: September 6, 2026

The Short Answer

The single most common mistake here is treating all three as the same kind of thing. They are not:

	EU AI Act Art. 50	California SB 942 / AB 853	C2PA
Type	Law	Law	Technical standard
In force	Aug 2, 2026	Aug 2, 2026	n/a (voluntary spec)
Visible label required	✅ Yes	Optional/manifest	❌ No
Machine-readable mark	✅ Where feasible	✅ Latent disclosure	✅ That’s the point
Public detection tool	❌ Not required	✅ Free and public	❌ No
Key deadline	Dec 2, 2026 (legacy systems)	In force now	—
Enforced by	EU market surveillance	California AG	Nobody

Two obligations, one mechanism. You comply with Article 50 and SB 942 using something like C2PA. C2PA on its own compels nothing and satisfies neither law completely.

EU AI Act Article 50 — The Visible-Plus-Invisible Rule

Applicable since August 2, 2026.

Article 50 imposes a two-limb transparency duty on generative AI:

Limb 1 — human-perceptible disclosure. AI-generated or AI-manipulated images, audio and video must carry a disclosure a person can actually notice: a visible label, an overlay, an audible notice. This applied from August 2, 2026 with no grace period.

Limb 2 — machine-readable marking. The same outputs must be marked in a machine-detectable way “where technically feasible,” so downstream systems can identify them as artificial or manipulated.

The split matters because the grace period only covers limb 2:

Generative systems already on the market before August 2, 2026 have until December 2, 2026 to implement machine-readable marking. Systems launched on or after August 2, 2026 had none.

⚠️ If you shipped a generative feature in, say, March 2026 and have not embedded provenance yet, December 2, 2026 is your date and it is inside the current quarter.

Territorial reach: Article 50 attaches to output reaching EU users, not to where your company sits. Non-compliance with the transparency obligations carries administrative fines under the Act’s penalty tiers.

California SB 942 / AB 853 — The Detection-Tool Rule

In force since August 2, 2026.

The California AI Transparency Act, as extended by AB 853, is narrower in scope but adds one obligation the EU does not:

1. A free, publicly accessible detection tool. Covered providers must let anyone check whether a piece of content was generated or substantially altered by their system. Not an enterprise API, not a paid tier — free and public. This is the requirement most teams discover late, because it is a product to build and host, not a flag to set in an encoder.

2. Latent disclosure. Every image, video and audio file generated or substantially altered by a covered system must carry an embedded, machine-readable provenance record.

California also runs a separate track on the input side: AB 2013, the Generative Artificial Intelligence Training Data Transparency Act, took effect January 1, 2026 and requires developers to publish a high-level summary of training datasets — sources and owners, number and types of data points, and whether the data includes copyrighted or personal information.

New York has moved in parallel with S 6954 on synthetic-content provenance and A 6578 on training-data transparency. At federal level the TRAIN Act, introduced January 2026, would give copyright holders a route to inspect training data. Treat the federal layer as pending, not operative.

C2PA — The Mechanism, Not the Mandate

C2PA Content Credentials is an open specification for a cryptographically signed manifest attached to a media file: what created it, when, what edits were applied, signed by a certificate that chains to a known issuer.

What it does well: tamper-evidence. If the manifest is intact, you can trust the chain. It is also the format the major platforms and camera makers have converged on, which makes it the pragmatic default.

⚠️ What it does not do: survive. A C2PA manifest is metadata. Screenshot the image, re-encode the video, strip EXIF — the manifest is gone and the file is now unmarked. The absence of a manifest proves nothing, because “no credential” and “credential removed” look identical.

That fragility is why the serious implementations layer:

Layer	Example	Survives re-encoding?
Signed manifest	C2PA Content Credentials	❌ No
Embedded watermark	SynthID (Google), vendor equivalents	✅ Mostly
Visible label	on-canvas badge, audible notice	✅ Unless cropped
Detection API	provider-hosted checker	✅ n/a

Article 50 limb 1 needs the visible label. Article 50 limb 2 and SB 942’s latent disclosure need the manifest and ideally the watermark. SB 942 additionally needs the detection endpoint. No single layer covers the set.

Which Applies to You

Situation	Article 50	SB 942	Build C2PA?
You generate images/audio/video, EU users	✅	maybe	✅
You generate media, California users at scale	maybe	✅	✅
You only generate text	limited	limited	optional
You host third-party generated media	as deployer	❌	helpful
You fine-tune and redistribute a model	✅	✅	✅

Note the text row. Both regimes focus their marking obligations on images, audio and video. Pure text generation carries different and generally lighter disclosure duties — chiefly telling people they are talking to a machine. Do not over-engineer provenance for a text-only chatbot; do not assume text is entirely out of scope either.

The Practical Read

If you ship generative media to any meaningful market in late 2026, build to the union of the two laws rather than picking one:

Visible disclosure on every generated image, video and audio output (Article 50, no grace period, already overdue if missing).
C2PA manifest embedded at generation time (both regimes).
A durable watermark underneath it, because manifests strip (both regimes, “technically feasible” reads badly if you skipped the robust option).
A free public detection page (SB 942 — this is the one with real lead time).
A published training-data summary if you are a developer, not just a deployer (AB 2013, live since January 1, 2026).

Then get items 2 and 3 live on legacy systems before December 2, 2026.

Last verified: September 6, 2026.