Skip to content
TheAgent Ecosystem
Models & Cost

AI Generated Content Labeling Requirement: Who It Binds

Two regimes took effect on 2 August 2026. One has a million-user threshold and the other has none, which inverts who is actually caught.

Muhammad Qasim HammadAI-assisted7 min read1,378 words

Updated November 5, 2026 — Corrected who the EU marking duty binds and the fine cap for SMEs.

AI-drafted, reviewed by Muhammad Qasim Hammad on November 3, 2026. See our AI disclosure.

Compliance: EU vs California
Table of contents
  1. What do the rules actually require?
  2. Why did both land on the same date?
  3. Does the California law apply to you?
  4. What is a latent disclosure and can you actually do it?
  5. What should you do before December?

Two labeling regimes took effect on the same day, and that was not a coincidence. The AI generated content labeling requirement now runs on both sides of the Atlantic from 2 August 2026, but one of them has a size threshold and the other does not. Which means the smaller you are, the more likely only the European rule reaches you.

What do the rules actually require?#

Marks a machine can read, and in California a tool the public can use. The EU requires providers of generative systems to mark output so software can detect it, and deployers to label deepfakes visibly. California requires covered providers to offer a free AI-detection tool plus visible and embedded disclosures.

Two kinds of disclosure are in play and the vocabulary differs by regime. A manifest disclosure is the visible one, a label a person can see on the content. A latent disclosure is embedded in the file itself, invisible to a reader and detectable by software, and it is the part with real engineering behind it.

The visible half is a product decision you can make this afternoon. The embedded half depends on what your generation stack actually supports, which for most teams means what their model vendor emits, since you are not going to write a provenance implementation yourself.

Four step explanation of manifest visible disclosure and latent embedded disclosure and where each one comes fromOne is a product decision you own. The other arrives from your vendor and can be destroyed downstream.

There is a third obligation people miss because it is phrased as an exception. Content generated and published before 2 August 2026 does not need retroactive labelling under the Commission's Article 50 guidelines, so the duty is prospective. That is genuine relief for anyone sitting on an archive of generated assets, and it means the work is bounded by what you publish from here.

California's framing is worth noting because it goes further than marking. Requiring a free public detection tool puts an obligation on the provider to let anybody check content against their system, which is a different kind of commitment from embedding a mark and hoping someone reads it.

Why did both land on the same date?#

Because California moved. The AI Transparency Act was originally due to take effect on 1 January 2026, and AB 853, signed on 13 October 2025, pushed it to 2 August 2026 to line up with the European provenance timeline. The alignment was deliberate rather than coincidental.

That matters for planning more than it sounds. A single 2 August date means one engineering effort can satisfy both regimes if the provenance approach is chosen sensibly, and it means the vendors you depend on were building to one deadline rather than two. It also means nobody gets a staggered runway.

EU, Article 50California, SB 942
Operative from2 August 20262 August 2026
Size thresholdNoneOver 1,000,000 monthly users
Visible labellingDeepfakes and public-interest textManifest disclosure on covered content
Embedded markingMachine-readable marksLatent disclosure where feasible
Public detection toolNot requiredRequired, and free
Penalty shapeUp to €15,000,000 or 3% turnover, whichever is lower for an SME$5,000 per violation, per day
Four figures covering the shared start date, the California user threshold, the per violation penalty and the European transitional deadlineThe threshold is the one that surprises people: it excludes small teams from California and not from the EU.

Read the threshold row twice, because it inverts the usual assumption. A team with 20,000 users is outside SB 942 entirely. Article 50 has no size threshold, so the European duty for the team's role still applies: marking if it provides a generative system, deepfake and public-interest labels if it deploys one. The rule that most builders assume is the distant one is the one that actually reaches them.

Does the California law apply to you?#

Only above a million. Covered providers are those that create or produce a generative AI system with over 1,000,000 monthly visitors or users that is publicly accessible in California. Below that line the Act's obligations do not attach, and most teams building products on someone else's model are comfortably below it.

The word doing the work is provider. SB 942 targets whoever produces the generative system, not everyone who uses one. If you are calling a vendor's API, the detection-tool obligation is theirs rather than yours, which is a different position from Article 50, where a deployer has labelling duties of its own for deepfakes and public-interest text, while the marking duty itself binds providers of generative systems.

Checklist of six common pipeline steps that strip embedded provenance metadata from generated mediaEvery item here happens after your vendor did its job correctly, and inside infrastructure you control.

Being outside SB 942 is not a reason to skip provenance. Crossing a million users is a growth outcome you would welcome, and retrofitting marks onto a generation pipeline after the fact is considerably harder than building it in while the volume is small.

What is a latent disclosure and can you actually do it?#

An embedded, machine-detectable mark that travels with the file rather than sitting next to it. California requires it where technically feasible, which is an unusually honest piece of drafting, because for a lot of content types it genuinely is not yet feasible and the standard bodies are still catching up.

Almost none of this is yours to implement. If you generate images or audio through a vendor, the provenance is applied at generation, and your job is to establish what your vendor emits and to avoid stripping it downstream. Pipelines that re-encode, resize or optimise media routinely destroy embedded metadata without anyone noticing.

Comparison of the European and Californian labeling regimes across thresholds, who is bound and what each requiresThe row that matters is the threshold. Below a million users only one of these columns applies to you.

That last failure is the one to check first, because it converts a compliant vendor output into a non-compliant published asset entirely inside your own infrastructure. An image optimiser that rewrites files on upload will happily discard whatever provenance arrived with them.

What should you do before December?#

Add the visible label now and use the remaining time on the embedded half. The European marking and detection obligation has a limited transitional period to 2 December 2026 for generative systems already on the market, and that is the only deferred piece in either regime.

Decision flowchart for content labeling duties, testing EU reach, the million user threshold and whether your pipeline preserves embedded marksMost small teams end at the same place: European duties apply, California does not, and the pipeline is the risk.

One planning caveat: the transitional period covers marking and detection, not the visible label and not the interaction disclosure. Reading it as a general extension is the mistake that will cost somebody a scramble in late November. Everything except embedded marking on already-live generative systems applied on 2 August.

Sequence it by cost. The visible disclosure is an afternoon, the vendor question is an email, and the pipeline audit is a morning spent checking whether your own processing strips what arrives. None of that requires a provenance implementation, and together it covers most of what either regime asks a small team to do. Whether your chatbot must announce itself is the interaction half of the same package.

Frequently asked questions

What does the AI generated content labeling requirement actually demand?
In the EU, that providers of generative systems mark output so it can be detected, and that deployers visibly label deepfakes. In California, that covered providers apply visible manifest disclosures, embed latent provenance disclosures where technically feasible, and offer a free public AI-detection tool. The visible half is a product decision; the embedded half depends on what your generation stack emits.
Does California's SB 942 apply to my company?
Only if you produce a generative AI system with more than 1,000,000 monthly visitors or users that is publicly accessible in California. Below that threshold the obligations do not attach. The word doing the work is provider, meaning whoever produces the system, so a team calling a vendor's API is generally outside it while the vendor is inside.
Why did the EU and California rules start on the same day?
Because California moved to match. SB 942 was originally due to take effect on 1 January 2026, and AB 853, signed on 13 October 2025, shifted it to 2 August 2026 to align with the European provenance timeline. The practical effect is that one engineering effort can satisfy both, and that vendors were building toward a single deadline.
What is the difference between manifest and latent disclosure?
A manifest disclosure is visible: a label on the content that a person can read. A latent disclosure is embedded in the file itself, invisible to a reader and detectable by software. The visible one is straightforward to add. The embedded one usually arrives from your model vendor at generation time, and your job is mainly to avoid destroying it.
Do we have to label content we published before August?
No. Under the European rules, content generated and published before 2 August 2026 does not require retroactive labelling, so the duty applies prospectively. Separately, a limited transitional period runs to 2 December 2026 for the marking and detection obligation on generative systems already on the market. That transition covers embedded marking only, not the visible label.

Sources

Primary references and vendor documentation used while drafting and reviewing this article.

  1. Jones Day: California Enacts AI Transparency Law Requiring Disclosures for AI Content
  2. European Commission: Commission starts enforcing AI Act rules and new transparency requirements on 2 August
  3. EUR-Lex: Consolidated EU AI Act, version of 27 July 2026
  4. European Commission: Guidelines on Article 50 transparency obligations, C(2026) 5054

Written by

Muhammad Qasim Hammad
Muhammad Qasim Hammad
AI agents & automationFounder · Cart Gaze LLCPMP-certified PM

Muhammad Qasim Hammad is an AI agent and automation expert and the founder of Cart Gaze LLC (cartgaze.com). He builds product for the love of it: when an idea lands, a working prototype is usually running within hours, built with the same AI agents and automations he sells. He puts his own output at roughly 20× what it was before agents, and the Agentic OS behind this site is the working proof, documented in public with the tools he actually ran and what they really cost.

AI & Automation Services

Want a pipeline like this running in your business?

I'm Qasim — I design and ship AI agents and n8n automations for solo operators and small teams. Tell me what's eating your team's week, and I'll scope a fix.

Related reading