Take one image produced by a generative model on a Tuesday afternoon. The file leaves the model, passes through an editing tool, gets posted by a marketing team, travels across a social platform, and lands on a phone screen. Six different parties touched it. Current law does not ask all six for the same signal. Law asks each of them for a disclosure in a different layer, on a different date, backed by a different penalty.
AI transparency requirements are layered, and that layering is the part most compliance programmes get wrong. Teams read Article 50 of the EU AI Act, conclude that AI output needs a label, and ship a watermark. Then a market surveillance authority asks which of the four statutory qualities that single technique satisfies, and whether the deployer who published the file added a human-perceivable disclosure on top of it. Two separate duties, two separate parties, one file.
The walk below goes outward, layer by layer. Innermost is the mark written into the bytes. Outermost is the camera sensor that will soon be obliged to sign what it records. Every layer below names the instrument, the obliged party, the date, and the record an auditor will ask for. If you are still working out which EU AI Act duties attach to your role, start there and come back, because Article 50 applies by role rather than by risk tier.
Article 50(2) obliges providers to write a machine-readable mark into the output itself
Article 50(2) governs the innermost layer, and the signal it demands is invisible. Providers of AI systems generating synthetic audio, image, video or text must mark the output in a machine-readable format and make it detectable as artificially generated or manipulated. The duty is a design duty. A provider satisfies it in the generation path rather than in the publishing workflow.
Article 50(2) leaves several categories outside the marking duty. Short strings of numbers, symbols or letters escape. Source code escapes. Machine-to-machine output escapes. Closed-loop industrial or product development output escapes unless the output is final. Standard assistive editing escapes, and so does processing that does not substantially alter the deployer’s input data or its semantics. The Commission’s own answers on these exclusions are the working reference, and they read narrowly.
Marking under Article 50(2) also carries four named qualities, set out in the reference table further down this page. Four adjectives sound like drafting filler until an authority asks you to evidence each one separately for the technique you chose.
Readiness against this transparency obligation has been measured. Researchers at Maastricht University examined 50 text-to-image systems and found machine-readable marking in 19 of them, or 38 per cent. Six of those 19 relied on metadata alone, which strips out of a file on routine upload. Three of the 40 systems outside the open-source part of the sample listed a provider established in an EU country, which tells you how much of the obliged population sits outside the Union and inside the Act’s reach anyway.
California’s latent disclosure carries six named fields inside the same layer
Inside the same innermost layer, California asks a different question. Rather than specifying qualities, the California AI Transparency Act specifies content. A latent disclosure must convey the covered provider’s name, the name and version information of the system that created or altered the content, the time and date of creation or alteration, a unique identifier, and whether the content was generated or modified by artificial intelligence. From 1 January 2029 a sixth field joins the latent disclosure: whether the system is designed primarily to function as assistive technology.
Article 50(2) and the California chapter diverge in drafting style, and the engineering consequence shows up fast. Article 50(2) is outcome-specified, so a provider must argue that its chosen technique meets four qualities. Chapter 25 of the California Business and Professions Code is field-specified, so a provider must demonstrate that six values are present and that the latent disclosure is permanent or extraordinarily difficult to remove or tamper with. One instrument tests the method. One instrument tests the payload. A single implementation can satisfy both transparency obligations, but only if it was designed against both lists rather than against a summary of either.
The covered provider population changed on 30 September 2026. Senate Bill 1000, signed on 30 September 2026 and effective at once, deleted the threshold of one million monthly users from the definition. A covered provider is now any person that creates, codes or otherwise produces a generative AI system publicly accessible in California. Small vendors who read the 2024 text and concluded the Act did not reach them were wrong eight weeks after it became operative.
The Code of Practice requires two marking layers when one cannot carry the four criteria
Above the raw mark sits a soft-law layer that has become the practical standard for AI transparency work. The Commission published its Code of Practice on marking and labelling AI-generated content on 10 June 2026, following a stakeholder process launched in November 2025. Adherence is voluntary. Consequence is not.
Here is the chain worth tracing. Signing the Code is an adequate route to demonstrating compliance with Article 50(2), Article 50(4) and Article 50(5), and the Commission and the AI Board have both assessed it as adequate. The Code states that where a single marking technique cannot on its own meet the four statutory qualities at the current state of the art, at least two layers of machine-readable marking are required. Metadata alone cannot survive a social upload, so metadata alone is a single technique that fails one of those four. Therefore a provider shipping metadata only has, by the Commission’s own reference document, an evidenced gap rather than an arguable position.
Two carve-outs from the two-layer expectation exist. Generative AI embedded in a closed, technically controlled physical product may run single-layer. Free-form text may run single-layer where it cannot carry metadata, and watermarking applies above 200 tokens, roughly 150 words, rather than to every short output. Providers signing the Code must also ban tampering in their terms or acceptable use policies, which closes the gap at the point of use. Signatories gain legal certainty irrespective of where they are established or supervised. Non-signatories must show adequacy by other means and should expect more information requests.
Uptake is measurable rather than theoretical. The Commission’s own record of who has signed listed 95 signatories to the provider section and 192 to the deployer section at its 24 September 2026 update, with roughly half described as small and recent companies. An organisation picks the provider section, the deployer section or both, and must accept every commitment in whichever section it picks. Cherry-picking single measures is not available.
Article 50(4) moves the visible label onto the deployer who publishes the content
Article 50(4) steps the signal outward, and the obliged party changes with it. Marking belongs to the provider. The human-perceivable disclosure under Article 50(4) belongs to the deployer. Deployers who generate or manipulate a deepfake must disclose that the content is artificially generated or manipulated. Deployers who publish AI-generated text to inform the public on matters of public interest must disclose that too.
Article 50(4) scope is narrower and stranger than teams expect. A deepfake under Article 3(60) needs three cumulative features: resemblance to existing persons, objects, places, entities or events; the existence of what is resembled; and a false appearance of authenticity. A deepfake forming part of an evidently artistic, creative, satirical or fictional work attracts a lighter duty, provided the disclosure does not interfere with enjoyment of the work. Deployers are exempt where the generation is authorised by law for law enforcement purposes.
The text limb of Article 50(4) turns on editorial responsibility. Human review means a person holds ultimate legal responsibility for publication, and the review has to be genuine rather than pro-forma. Spell-checking does not count. A newsroom with a named responsible editor is outside the labelling duty on that output. A content operation that passes drafts through a reviewer with no publication responsibility is inside it.
One point decides whether a deployer has satisfied this transparency obligation at all. A provider’s machine-readable mark does not discharge the deployer’s duty. Disclosure must be clear and distinguishable, delivered at first exposure at the latest, and comply with applicable accessibility requirements, which Article 50(5) points to rather than creates. The Commission has published three optional icons for the purpose: a basic AI-content mark, one for fully AI-generated content, and one for content partially modified with AI. Teams running chat assistants and retrieval pipelines should map this against how they already govern chatbots and retrieval systems in production, because the disclosure surface is the product surface.
SB 1000 deleted the manifest disclosure that SB 942 had made mandatory
Something unusual happened in this layer in California, and it is worth naming because it inverts the usual direction of travel. Senate Bill 942, as amended by Assembly Bill 853, required covered providers to offer users the option of a manifest disclosure, meaning a visible mark on image, video or audio content. Chapter 25 became operative on 2 August 2026. Senate Bill 1000 removed that requirement on 30 September 2026, as an urgency statute taking immediate effect.
Fifty-nine days separated the duty becoming operative and the duty being deleted. California now regulates the invisible layer through the latent disclosure and leaves the visible layer to the deployer, which converges on the European split. Practitioners who built a user-facing toggle in July were building against a requirement that no longer exists, while the latent disclosure they treated as the optional half became the whole obligation.
Protection of the mark moved the other way. California bars platforms from knowingly stripping provenance data from content, from 1 January 2027, which pairs with the Code obligation on providers to prevent users doing the same thing at source. A mark protected at both ends is only as good as the inventory that says which systems produce it, so the register of systems in scope is the first artefact to get right.
Penalties moved in the opposite direction too. The baseline civil penalty remains 5,000 dollars per violation, collectable by the Attorney General, a city attorney or a county counsel, with each day of violation counting separately. Senate Bill 1000 added a 50,000 dollar penalty for falsely representing a system as qualifying for the temporary assistive technology exemption, repealing on 1 January 2029. The cheaper duty went away. The expensive misrepresentation arrived.
Article 50(1) requires the interface to name both the AI system and the party behind it
Article 50(1) governs the next layer out, which is the conversation itself. Providers must design systems that interact directly with people so that those people are informed they are dealing with an AI system, unless that fact is obvious to a reasonably well-informed, observant and circumspect person. The exception is read narrowly, and the test is cumulative on four points: the thing is an AI system, it enables a two-way exchange, the AI communicates directly, and the counterpart is a natural person.
Agents raised the bar. The final Guidelines adopted on 20 July 2026 state that an AI agent must disclose its AI nature and the person or entity on whose behalf it acts. Burying the fact in terms of service does not satisfy that duty. A generic label such as “assistant” does not satisfy it. Metadata alone does not satisfy it. Where interaction with a person is possible, the agent discloses itself by design, including at key operational steps rather than once at the start.
Guidelines on AI transparency are not binding. Market surveillance authorities will nonetheless treat them as the primary interpretive tool, which makes them binding in the only sense that matters during an inspection. Anyone shipping autonomous agents should read the disclosure duty alongside the controls that constrain an agent’s tools and human review, since both end up evidenced from the same logs.
Large online platforms inherit a surfacing duty on 1 January 2027
Beyond the publisher sits the distributor, and California is the jurisdiction that regulates it directly. Assembly Bill 853 added transparency obligations to platforms rather than to model builders, and Assembly Bill 2713 rewrote the platform section before it took effect. Large online platforms, meaning public-facing social media, file-sharing, mass messaging and standalone search services above two million unique monthly users, acquire their obligation on 1 January 2027.
Three duties land on those platforms at once. A platform must detect provenance data in uploaded content, give users an interface showing whether content was generated or altered by a generative system, and let users inspect the system provenance data itself. A platform must also not knowingly strip system provenance data or a digital signature from content uploaded, distributed or downloaded, so far as that is technically feasible. GenAI hosting platforms acquire a separate duty on the same date, and may not knowingly make available a system that fails to place the required disclosures.
One carve-out in that section decides how much of it bites. A platform owes nothing in respect of provenance data that is not compliant or interoperable with widely adopted specifications from an established standards-setting body. Read that against the Code’s 2 February 2027 interoperability date and the consequence follows: a mark built outside recognised specifications is a mark California excuses platforms from surfacing, which makes interoperability a distribution question rather than a technical preference.
Senate Bill 1000 left those dates alone. A provider therefore faces a sequencing problem that is easy to miss. Marking duties bite now. Platform surfacing duties bite in January. A mark that no distributor reads produces no user-visible outcome in 2026 and a compliance dependency in 2027, because the platform duty assumes the provenance data is already in the file. Teams tracking the wider pattern across jurisdictions will find the state-by-state picture in the United States moves faster than the federal one.
Capture device manufacturers must embed provenance by default from 1 January 2028
The outermost layer is hardware, and the date is 1 January 2028. Section 22757.3.3 requires capture device manufacturers to offer users a latent disclosure option and to embed latent disclosures by default in devices first produced for sale in the state on or after that date. Cameras, phones and recorders become provenance signers.
The payload at this layer is lighter than the generative one. A capture device latent disclosure carries three fields rather than six: the manufacturer’s name, the device name and version, and the time and date of creation or alteration. Compliance is owed only so far as it is technically feasible and consistent with widely adopted specifications from an established standards-setting body, which is the same carve-out the platform section uses.
Follow the logic to its conclusion. If authentic capture is signed at the sensor and synthetic generation is marked under Article 50(2) at the model, then unsigned content becomes the anomaly rather than the default. Inverting the default is the long-run design intent behind both regimes, which explains why marking is specified as a durability property rather than a labelling convention. Manufacturers building for a 2028 ship date are designing to it in 2026.
Detection access decides whether a mark qualifies as machine-readable at all
A mark nobody can read is not machine-readable in any useful sense, which is why the Code treats detection as part of the marking duty rather than a courtesy. Signatories may run a shared detection solution jointly. Interoperability across detection is expected by 2 February 2027. Providers below one million monthly users may charge a reasonable fee where a single requester’s volume is unreasonable and detection carries substantial cost, but specified authorities, media organisations, fact-checkers, trusted flaggers, researchers and civil society organisations get access without charge.
Scale on the detection side has arrived faster than scale on the legal side. Google reported on 7 October 2026 that it had watermarked over 180 billion images and videos plus 240,000 years of audio, opened its detector globally in English, and was handling more than a million verification requests a day through Search, Gemini and Chrome. Partner coverage extends to several other model providers.
A caution belongs with that figure. Any detector verifies the watermarks it knows. Content from an unaffiliated model returns no result, and no result is not evidence of human authorship. The same Maastricht sample found visible disclosure on only 9 of 50 systems, and in 5 of those 9 the user could switch the marking off with a checkbox or remove it by upgrading a plan. Marking a user can disable fails the durability quality, which returns the question to the four criteria where it started.
Market surveillance authorities sample records rather than read intentions
Enforcement of Article 50 sits with national market surveillance authorities. The AI Office holds a narrower role, limited to systems built on general-purpose models where the same entity provides both, and to systems integrated into services designated as very large online platforms or search engines under the Digital Services Act. The European Data Protection Supervisor covers AI used by Union institutions. Ceilings reach 15 million euro or 3 per cent of total worldwide annual turnover for the preceding financial year, with Article 100(3) setting 750,000 euro for Union institutions and proportionality available to small and medium enterprises and small mid-caps. The detail of how the Act sets its fine tiers is worth reading before budgeting for the risk.
What gets sampled is narrower than the duty and broader than the code. An inspector asks which systems in your estate generate synthetic output, which marking technique each one applies, which of the four qualities that technique evidences, what the detection route is, whether the deployer-side disclosure under Article 50(4) fires at first exposure, and when each decision was taken and by whom. Answering means holding a per-system record rather than a policy statement, and the record has to predate the question.
AI transparency work fails inside otherwise mature programmes for a dull practical reason. Marking is an engineering decision logged in a repository. Disclosure is a product decision logged in a design review. Adequacy is a legal position logged nowhere by default. govern365.ai exists to put those three in one place, so the per-system marking choice, the deployer disclosure configuration and the adequacy rationale sit against the same system record and carry the same timestamps. Teams approaching this from scratch usually start from what to record about each model version, then widen the record to cover output handling.
| Layer | Obliged party | Instrument | Live from |
| Machine-readable mark in the file, which must be effective, interoperable, robust and reliable so far as technically feasible | Provider | AI Act Article 50(2) | 2 August 2026, or 2 December 2026 for prior systems |
| Latent disclosure fields | Covered provider | California BPC 22757.3 | 2 August 2026 |
| Visible disclosure of deepfakes and public-interest text | Deployer | AI Act Article 50(4) | 2 August 2026 |
| Notice that the counterpart is an AI system | Provider | AI Act Article 50(1) | 2 August 2026 |
| Surfacing of provenance data | Large online and hosting platforms | California BPC 22757.3.1 and 22757.3.2 | 1 January 2027 |
| Detection interoperability | Code signatories | Transparency Code of Practice | 2 February 2027 |
| Provenance signed at capture | Capture device manufacturers | California BPC 22757.3.3 | 1 January 2028 |
Four dates govern the transparency transition, and one falls in December 2026
Sequencing is where most remediation plans break, so take the dates in order. Article 50 has applied since 2 August 2026, and the California chapter became operative the same day. Regulation (EU) 2026/1744, adopted on 8 July 2026, left Article 50 in force. The same regulation moved the Annex III high-risk obligations to 2 December 2027 and the Annex I product obligations out to August 2028. AI transparency was not delayed with the rest.
One concession did arrive in that regulation. A new Article 111(4) gives providers of generative systems placed on the market before 2 August 2026 until 2 December 2026 to take the necessary steps to comply with Article 50(2). The regulation entered into force on 27 July 2026, and the concession covers the marking duty and nothing else. Systems placed on the market from 2 August 2026 onward never had it. A new Article 5 prohibition on systems that generate child sexual abuse material or non-consensual intimate material carries the same 2 December 2026 date, so one diary entry covers both. Recital 38 describes the marking concession as a transitional period of four months, which 2 August to 2 December matches exactly. Content created before 2 August 2026 needs no retroactive labelling, though the Commission encourages it.
So the live question on 24 October 2026 is a narrow one. If your generative system shipped before August, you have until 2 December to produce marking an authority would accept against all four qualities, and you have the Code’s two-layer expectation as the benchmark against which a single-layer answer will be read. If it shipped after August, the duty was never suspended, and the gap is already a gap. United States companies working out whether Article 50 reaches them should start with when a United States company is caught by the Act.
Two further dates close the sequence. Detection interoperability is expected by 2 February 2027, and platform surfacing duties in California begin on 1 January 2027, which means the mark you ship in December gets read by a third party a month later. Building the marking decision, the disclosure configuration and the adequacy argument as one auditable record rather than three disconnected artefacts is the difference between a December deadline met and a December deadline reported. Firms that want that record maintained rather than reconstructed can see how govern365.ai runs EU AI Act compliance as a continuous evidence process.
One last habit worth adopting. Write the tense into the record. Article 50 applies, the grace period ends, the manifest disclosure was required and no longer is. A register that records duties without recording when each sentence stopped being true is the register that produces a confident wrong answer during an inspection, and keeping those records somewhere an auditor can sample them is the cheapest part of the whole exercise. Where AI transparency is one regime among several, the same record serves all of them, which is the premise govern365.ai is built on.
Questions compliance teams ask about AI transparency
When did the EU AI Act transparency obligations start to apply?
Article 50 has applied since 2 August 2026. The Digital Omnibus delayed the high-risk regime but left the transparency rules in force on their original date. Providers of generative systems placed on the market before 2 August 2026 have until 2 December 2026 to comply with the Article 50(2) marking duty, under the new Article 111(4).
Is one watermark enough to satisfy Article 50(2)?
Usually not. Marking has to clear four statutory qualities so far as technically feasible, and the Code of Practice states that where a single technique cannot clear all four at the current state of the art, at least two layers of machine-readable marking are needed. Single-layer marking is accepted for free-form text, which cannot carry metadata, and for generative AI embedded in closed technically controlled physical products.
Who must disclose a deepfake, the provider or the deployer?
The deployer. Article 50(2) puts the machine-readable mark on the provider, while Article 50(4) puts the human-perceivable disclosure on whoever generates or disseminates the content. A provider’s mark does not discharge the deployer’s duty, which has to be clear, distinguishable and delivered at first exposure at the latest.
Do chatbots and AI agents have to say they are AI?
Yes, under Article 50(1), unless the fact is obvious to a reasonably well-informed and observant person, and that exception is read narrowly. The Commission’s Guidelines go further for agents, requiring disclosure of both the AI nature and the party the agent acts for. A reference in terms of service, a vague label such as “assistant”, or metadata alone will not do.
Does the California AI Transparency Act still require a visible AI label?
No. Senate Bill 942, as amended by Assembly Bill 853, required covered providers to offer users a manifest disclosure, and the chapter became operative on 2 August 2026. Senate Bill 1000, signed on 30 September 2026 as an urgency statute, deleted that requirement 59 days later. The latent disclosure embedded in the content remains mandatory.
Which companies count as covered providers in California?
Any person that creates, codes or otherwise produces a generative AI system publicly accessible in California. Senate Bill 1000 deleted the former threshold of one million monthly visitors or users, so vendors who sat outside the definition in August 2026 sat inside it by October.
Does AI-generated text always need a label?
No. Article 50(4) reaches text published to inform the public on matters of public interest, such as politics, public health, public security and environmental protection. Text that has undergone genuine human review or editorial control is exempt, where a named person holds ultimate legal responsibility for publication. Spell-checking is not review.
What are the penalties for getting AI transparency wrong?
In the Union, up to 15 million euro or 3 per cent of total worldwide annual turnover for the preceding financial year, whichever is higher, with Article 100(3) setting 750,000 euro for Union institutions. In California, 5,000 dollars per violation with each day counted separately, rising to 50,000 dollars for falsely claiming the assistive technology exemption.
