The Editorial Test: How an August 2 Deadline, a 190-Signature Code, and a Spell-Check Carve-Out Are Deciding Who Has to Say “It Was AI”
On August 2, 2026, the EU AI Act’s transparency obligations became the first binding content-labeling regime to take effect in a major Western market. The rules do not ban AI in news. They put a price on publishing machine-written text that nobody reviewed.
What Changed on August 2
The EU AI Act entered into force on August 1, 2024, and applies in stages. The stage that matters for media arrived on August 2, 2026, when Article 50 — the Act’s transparency article — took effect. From that date, four obligations bind the parties involved in producing and publishing synthetic content:[1]
- Chatbots and AI agents must announce themselves (Article 50(1)). Providers of systems that interact directly with people must tell users they are dealing with AI, unless it is obvious to a reasonably well-informed user. The guidelines are explicit that disclosure buried in terms and conditions, readable only by machines, or stated generically (“this system uses LLMs”) is not sufficient — and AI agents must disclose both their artificial nature and the person on whose behalf they are acting.[2]
- Providers must mark machine-generated outputs machine-readably (Article 50(2)). The company that builds a generative system must ensure its outputs carry marks a computer can detect as artificially generated — the invisible layer underneath the visible label.
- Emotion recognition and biometric categorisation must be disclosed (Article 50(3)), at the latest on first exposure to the people being scanned.[1]
- Deepfakes and AI-generated public-interest text must be labelled (Article 50(4)). Whoever publishes image, audio, or video content that “resemble[s] existing persons, objects, places, entities, or events” — or text published to inform the public on matters of public interest — must disclose its artificial origin, subject to one exception.[1]
Enforcement sits with national market surveillance authorities, the European Commission’s AI Office, and, where EU institutions are the culprits, the European Data Protection Supervisor. Companies face fines of up to 15 million euros or 3% of global annual turnover, whichever is higher; EU institutions face up to 750,000 euros, with proportionality rules tempering the numbers for small and medium-sized enterprises.[1]
The design logic is worth pausing on, because it explains why this landed in the media domain and not, say, the consumer-safety domain. The Commission’s own framing is that AI has made it “increasingly difficult to distinguish AI-generated and manipulated content from human-created and authentic content,” creating “new risks of misinformation and manipulation at scale.”[1] In other words, the reader-side test — can a person tell? — has failed. So Article 50 moves the detection burden upstream from the audience, who cannot tell, to the publishers and providers, who can. The law does not require anyone to stop using AI. It requires the people who know what the machine did to say so.
Two scoping decisions define the regime’s practical reach. First, the deepfake test is objective: intent to deceive is not required, and photorealism is not decisive — the content need only depict a subject that is “plausibly real,” which catches AI-generated actors and digital replicas while leaving wholly fantastical content outside.[2] Second, the public-interest trigger for text is construed broadly, covering “politics and democratic process, public services, justice, fundamental rights, public security, health, environment, consumer safety, and economic, financial, scientific and cultural matters relevant to public debate” — but not AI-generated fiction, product descriptions, or a chatbot reply seen only by the person who asked.[2] A few technical deadlines were also deferred under the Digital Omnibus revisions: systems already on the market have until December 2, 2026 to meet the marking obligation, and watermark-detection interoperability is required by February 2, 2027. Content published before August 2, 2026 does not need to be labelled retroactively.[2]
The Spell-Check Line: How the Editorial Exception Works
Article 50(4) is the provision that reaches newsrooms, and its entire regulatory weight sits on one exception. AI-generated public-interest text must be disclosed unless “the AI-generated text has undergone a process of human review or editorial control by a natural or legal person with responsibility.”[2] That carve-out is where the Commission drew the line between journalism and automation — and the finalised guidelines, published July 20, 2026, make the bar deliberately high:
- Competence is required. The reviewing person must have the relevant competence to judge the substance; the review must be a real examination, not a glance.[2]
- Spell-checking does not count. The guidelines state plainly that “spell-checking or cursory sign-off” is not sufficient to trigger the exception.[2]
- The exception can be lost after the fact. Any substantive AI edit applied after the human review removes the exception — the review has to be the last meaningful pass.[2]
- Assertions are not enough. Under the Code of Practice, claiming the editorial exception “requires documented workflows, not solely assertions that such review took place.”[2]
The Media and Journalism Research Center, which is running a multi-country study of how European news organisations are implementing the rule, captures what this design actually does:
“It does not ban newsrooms from using AI, and it does not require a label on everything AI touches. It ties the duty to disclose to how editorial oversight and responsibility are organised. That makes it a question about how journalism is run, not about technology.”[3]
That is the core logic of the whole regime. The EU did not write a technology rule that asks “was AI used?” — a question with a binary, often meaningless answer, since nearly every modern newsroom uses AI somewhere in its pipeline. It wrote an accountability rule that asks “who stood behind what was published?” A newsroom with a real editor is regulated by its own editor. A pipeline that auto-publishes machine text is regulated by the label it must now carry. Labrador, a CMS vendor that built Article 50 compliance tooling for newsrooms, puts the dividing line in one sentence: “publishers who keep a human in the loop should feel very little of Article 50. Publishers who auto-publish AI text carry all of it.”[4]
There is a catch that cuts both ways, and CMS vendors are already building for it: under a regime where the exception requires documented workflows, “you may someday need to prove the human review happened, not just that it could have.”[4] The practical consequence is that the edit log becomes a compliance artifact — the same workflow feature that was once an internal newsroom nicety is now the evidentiary basis for the entire exemption.
The 190-Signature Code: Regulating by Codebook Instead of Courtroom
Article 50 tells publishers what to do; the Code of Practice on Transparency of AI-generated Content tells them how, and it is the most quietly consequential document in this story. Drafted by independent experts, finalised in June 2026, and assessed as adequate by both the Commission and the AI Board, the Code offers signatories “a streamlined, predictable and legally certain pathway to guarantee and demonstrate compliance, regardless of the place where the organisations operate or the competent supervisory authority.”[5] By end of July 2026, about 190 organisations had signed; as of the September 17, 2026 update, the Code counts 95 signatories to its provider section and 192 to its deployer section, with about half being small and recent companies.[5]
The signatory list reads as a who’s-who of the generative stack: Anthropic, Google, Meta, Microsoft, Mistral, OpenAI, Black Forest Labs, Cohere, and Synthesia headline the provider section; Getty Images, Bulgari, Lenovo, Lufthansa, and Iberdrola appear among deployers.[5] Voluntary in name, important in practice — because the alternative to signing is proving your own measures are adequate, authority by authority, across twenty-seven member states.
The Code’s most honest finding is technical. The Commission “has not identified any single marking or labelling technique that currently meets the AI Act’s standards,” meaning a multi-layered approach is the working requirement.[2] Concretely, the provider commitments include:
- At least two machine-readable marking techniques for most outputs — for example, digitally signed, tamper-evident metadata plus imperceptible watermarking — with an obligation to prohibit users from stripping metadata. A single technique suffices only for free-form text (which cannot carry metadata) and for systems built into closed physical products.[2]
- A short-text exemption with a sunset. Text under 200 tokens — roughly 150 words — is exempt from the watermarking requirement, though the Code notes the threshold may decrease as watermarking advances.[2]
- A free detection mechanism. Providers must let third parties verify, with confidence scores, whether content came from their system — free of charge, with regulators, law enforcement, media, fact-checkers, researchers, and civil society always getting unrestricted access.[2]
- Interoperability by February 2, 2027, so content can be verified without running each provider’s detector in turn.[2]
On the deployer side, labels must be “immediately recognisable without user interaction, remain visible for a sufficient duration, be embedded directly into the content where possible, and be perceivable at the time of first exposure.”[2] The Commission has finalised three EU icons — for fully AI-generated content, for AI-modified content, and a basic icon with an interactive second layer — plus audible disclaimers for audio-only content where no visual label is possible.[2] The unifying principle across both sections is the same one that drives the editorial exception: disclosure has to survive the moment of consumption. A label the reader never sees, or a mark that disappears when a file is copied, is treated as no disclosure at all.
Signatories will also be invited into two task forces launching in September 2026, tasked with sharing implementation practices and advancing the state of the art.[5] That is the Code’s real function: it converts a statutory obligation into a running technical conversation, so that the standard evolves with the technology instead of being relitigated in each enforcement case.
What It Means for Newsrooms
For working journalists, the rule’s effect is narrower than the headlines suggest, and larger than the letter of the law. The Media and Journalism Research Center’s implementation guide distills Article 50(4) into three sequential questions, each of which must be answered before the next is reached:[3]
- Did an AI system generate or manipulate this text, used by the outlet under its own authority? A journalist writing unaided, or a tool used outside the organisation’s control, is outside the provision.
- Is it published to inform the public on a matter of public interest? The Commission reads this broadly, but routine commercial text is out.
- Did it have real human review or editorial control, and does a person or organisation hold editorial responsibility for publishing it? Review means substantive examination by someone with the judgement to do it; editorial control means authority to approve, change, or reject; editorial responsibility means someone ultimately carries legal responsibility. Fixing spelling does not count.[3]
Only when all three conditions line up — AI under the outlet’s control, public-interest purpose, no qualifying human review — does the disclosure duty bite. In practice, newsrooms that already operate like newsrooms are largely covered: Labrador’s scenario analysis concludes that no label is needed for an AI draft substantively edited by a journalist, for AI titles and summaries edited by a journalist, or for assistive tooling like spellcheck and light retouching.[4] The outlets whose workflows actually change are the ones running automated publishing — wire-style generated coverage, SEO farms passing as news, synthetic explainers with no editor in the loop. The law’s burden lands hardest where the editorial layer is thinnest.
The MJRC’s research design also makes an honest point that deserves to travel beyond academia: from the outside, you cannot tell compliance from silence. An article with no disclosure may simply have been written by a person, fall outside the rule, or fall within the exception — so the study commits to recording “that no disclosure was observed in the material we sampled” and never writing that AI material was unlabelled, “unless the outlet has said the material was AI generated.”[3] That is the measurement problem enforcement authorities face too, and it is why the Code leans so heavily on machine-readable marking: when visibility fails, verifiability has to carry the weight.
What to Watch
Four dates and one open question will decide whether this regime becomes the global default or an EU-only curiosity. The task forces launch in September 2026, where the marking standards will be refined in practice.[5] The MJRC’s first report on how newsrooms are actually implementing the rule lands in November 2026 — the first systematic look at whether outlets document editorial review or quietly relabel automation as editing.[3] December 2, 2026 closes the deferral for generative systems already on the market, and February 2, 2027 forces watermark-detection interoperability.[2] Law firms read the enforcement window the same way: as Cooley’s compliance alert puts it, “noncompliance can trigger fines of up to €15 million or 3% of worldwide annual turnover, whichever is higher.”[6]
The open question is the direction of the editorial exception itself. It could become a floor — a documented, audited standard for what counts as editorial review that pushes every automated outlet to add humans back into the loop. Or it could become a loophole — a paperwork exercise where the “documented workflow” exists to satisfy a regulator but the machine still writes the news. The first enforcement actions against auto-publishers, and the first newsroom that loses the exception because a substantive AI edit happened after the editor signed off, will show which way the incentive really points. What is already settled is the principle: in the EU, the question “who wrote this?” now has a legal answer, and someone has to be able to prove it.
References
[1] European Commission — [PRIMARY] “Safer and more transparent AI.” Directorate-General for Communication. August 2, 2026.
[2] Paul, Weiss, Rifkind, Wharton & Garrison LLP — [SECONDARY] “EU Finalises Transparency Rules for AI-Generated Content.” August 4, 2026.
[3] Media and Journalism Research Center — [SECONDARY] “Who Has to Say It Was AI? Transparency and Editorial Responsibility under Article 50(4) of the EU AI Act.” 2026.
[4] Labrador CMS — [SECONDARY] “A Newsroom’s Survival Guide to the EU AI Act’s Article 50 Transparency Rules.” 2026.
[5] European Commission, AI Office — [PRIMARY] “Strong Backing for the Code of Practice on Transparency of AI-Generated Content.” July 31, 2026; updated September 17, 2026.
[6] Cooley LLP — [SECONDARY] “EU AI Act: Transparency Obligations Take Effect 2 August 2026.” August 2026.
AI automated and generated content. Human reviewed.
Listen to the Bears Discuss This

Episode 28 of AI Policy for the Bears