AI disclosure rules 2026: EU AI Act and California SB 942

If you publish an edited Gemini image or Veo clip into the EU or California, two disclosure laws now share one start date. The EU AI Act's Article 50 and California's SB 942, as amended by AB 853, both take effect on 2 August 2026. Neither one cares whether you cropped out a corner watermark first — cleaning up pixels and satisfying a disclosure law are different acts entirely.

Why 2 August 2026 matters

Before AB 853, California's SB 942 was on a separate operative date: 1 January 2026. AB 853 pushed that date to 2 August 2026, landing it on the same day Article 50 of the EU AI Act starts applying. A business shipping into both markets now faces one deadline instead of two, and a solo creator posting to an EU marketplace or a California-headquartered platform inherits both timelines at once, regardless of where they personally live.

Neither regime is new law waiting to be written. Both were already on the books; 2 August 2026 is when their transparency obligations start biting.

Providers, deployers, and the distinction most guides skip

Article 50 does not impose one flat obligation on everyone who touches an AI image. It splits duties by role, and almost every ranking page on this topic collapses that split into a single "you must label AI content" line. That collapse is where the confusion starts.

• Providers — the companies that build and supply the AI system, like Google for Gemini — must mark synthetic output in a machine-readable format so it can be detected as artificially generated (Article 50(2)). This is the SynthID layer: invisible, embedded at generation time, Google's responsibility.
• Deployers — anyone who uses the system professionally or publicly — carry a narrower duty under Article 50(4): if the output is a "deep fake," resembling a real, existing person, place, or event closely enough to falsely appear authentic, disclose it. The same clause covers AI-generated text on a matter of public interest, unless it went through documented editorial review.
• Article 50(4) narrows around evidently artistic, creative, satirical, or fictional work — disclosure there only needs to be appropriate and non-intrusive, not a blaring label.

Read literally, a fictional AI product shot — a water bottle on a rock, a bag on a table, nobody's face in frame — usually does not meet the Article 50(4) "deep fake" definition, because it does not resemble an existing person, place, or event. The provider-side marking duty still applies to that image, but it applies to Google, invisibly. Put a recognizable real person, a real location, or a real brand context into the frame and presented as authentic, and the deployer duty in Article 50(4) is squarely aimed at you.

What the EU AI Act requires

Article 50's marking obligation applies to providers from 2 August 2026, with one exception: systems already on the market before that date get a grace period on the marking mechanism specifically, running to 2 December 2026. That runway does not delay the deployer duties in Article 50(4), which apply on 2 August 2026 regardless of when the system launched. Non-compliance can draw fines up to €15 million or 3% of worldwide annual turnover, whichever is higher. None of this retroactively requires labeling content generated before 2 August 2026.

What California SB 942 requires

SB 942 is built around two disclosure types:

• Manifest disclosure is a visible marker on the content itself — clear, conspicuous, and understandable to an ordinary viewer, permanent or hard to strip where technically feasible.
• Latent disclosure is an embedded, non-visible signal carrying provenance data: the provider's name, the system name and version, a creation or alteration timestamp, and a unique identifier, detectable through a free public AI-detection tool.

SB 942's direct obligations fall on "covered providers" — generative AI systems with more than one million monthly visitors that are publicly accessible. That is Google, not a freelancer posting a Gemini-generated image to a portfolio site; your own SB 942 exposure as an individual publisher is close to zero under the current provider-facing text. What does reach you is the platform layer: starting 1 January 2027, large online platforms and GenAI hosting platforms pick up duties to preserve provenance data and disclose AI or device origin — the places you post to may start asking more of every upload, even though the law asks nothing of you directly today.

Removing a corner watermark does not touch what either law is marking

This is the point every generic "AI watermark law" roundup glosses over. Article 50's machine-readable mark and SB 942's latent disclosure are both invisible, embedded signals — SynthID and comparable provenance metadata — layered in at generation time by the provider. The visible Gemini logo in the corner of an image is a separate thing: Google's own branding element, not the legal marking mechanism either statute regulates.

This project's browser eraser works on that visible corner logo only. It does not remove, weaken, or interact with SynthID or any latent disclosure metadata, and it is architecturally incapable of doing so — SynthID is a distributed, learned signal, not a fixed pixel pattern a detector can patch. Cleaning up the visible mark changes what a viewer sees in the corner; it changes nothing about whether Article 50 or SB 942 applies to your publication. See visible Gemini watermark vs SynthID for how the signals differ, and what this tool cannot remove for the honest boundary of pixel editing.

Should you disclose this image? Run the checklist

Do you need an AI disclosure for this file?

Work through these in order. A 'stop' outcome means the checklist doesn't apply yet — recheck it the moment that changes.

1. 1Does the image resemble a real, identifiable person, place, or event closely enough that it could pass as authentic? — Yes

   Treat it as a likely Article 50(4) deep fake. Add a clear, visible AI-generated disclosure before you publish, unless the piece is evidently satirical or artistic, or went through documented editorial review.

2. 2Is this AI-generated text on a public-interest topic — news, elections, health, or similar — published without independent human review? — Yes

   The Article 50(4) text-disclosure duty likely applies. Label the text as AI-generated, clearly, at first exposure to a reader.

3. 3Will this reach an EU audience or a California resident, or are you unsure where your audience is? — Yes or unsure

   Default to disclosing. A visible AI label costs little, and there is no upside to guessing wrong about your audience's location.

4. 4Is this a private file you are not publishing, selling, or handing to a client? — Yes

   Stop here. Neither regime targets undistributed private use. Re-run this checklist the moment the file leaves your hands for any public or commercial purpose.

5. 5Are you removing the visible Gemini corner mark before publishing? — Doesn't change the answer

   That edit is irrelevant to this checklist. Removing a visible logo does not create or discharge a disclosure duty. Decide your disclosure independently of what the corner of the image looks like.

Frequently asked questions

When do the EU AI Act and California SB 942 disclosure rules take effect?

Both land on the same date. The EU AI Act's Article 50 transparency duties apply from 2 August 2026, and AB 853 moved California SB 942's operative date to 2 August 2026 specifically to align with it. Before AB 853, SB 942 had been set for 1 January 2026.

What does the EU AI Act require for AI-generated content?

Providers must mark synthetic output in a machine-readable format under Article 50(2), with a grace period on that marking mechanism running to 2 December 2026 for systems already on the market before 2 August 2026. Separately, deployers publishing a "deep fake" — content resembling a real person, place, or event closely enough to appear authentic — must disclose that under Article 50(4), with no grace period. Non-compliance can draw fines up to €15 million or 3% of worldwide annual turnover.

Does removing a visible Gemini watermark satisfy or violate disclosure law?

Neither, automatically. Removing a visible mark changes pixels, not your legal disclosure duty. The mark Article 50 and SB 942 actually regulate is an invisible, embedded signal — SynthID and comparable metadata — that this project's eraser does not touch and cannot touch. Whether a use requires a manifest disclosure, a latent signal, or both depends on jurisdiction, audience, and platform. This is not legal advice.

When do hosting platforms get separate SB 942 duties?

SB 942 adds hosting-platform duties starting 1 January 2027 — for large online platforms and GenAI hosting platforms — later than the manifest and latent disclosure duties that apply to covered providers from 2 August 2026. Confirm current requirements against the bill text before relying on any single date.

Where to go next

• Whether editing out the mark is itself permitted: is it legal to remove a Gemini watermark.
• The practical ethics checklist: how to remove a Gemini watermark responsibly or the responsible-use summary.
• How a visible logo, SynthID, and Content Credentials differ: visible Gemini watermark vs SynthID.
• Video's own labeling reality: Veo watermarks, YouTube AI labels, and monetization.
• The honest edge of pixel-level processing: what this tool cannot remove, then the browser eraser.