If your content reaches EU readers and addresses a matter of public interest, Article 50 of the EU AI Act requires you to label it as AI-generated. Noncompliance now results in penalties of up to €15 million or three percent of worldwide annual global turnover for the preceding financial year, whichever is higher.
There is a way around this, however. When content goes through a genuine human review, with someone holding editorial responsibility, the AI label isn’t necessary. The catch is that Article 50 sets a higher bar for human reviews than what most teams provide, and AI editing checklists that many teams use today don’t meet the Article’s standards.
This piece breaks down what the law specifically asks for, where teams’ current processes fall short, and a four-part test you can run against your own workflow to ensure compliance.
Article 50 covers four separate situations:
This post focuses on the latter. If your organization uses AI to draft or rework text and publishes it to inform readers on a public interest matter, you must disclose that the text is AI-generated. This obligation took effect on August 2, 2026, alongside most of the AI Act’s other provisions, and it applies regardless of where you register your company. What matters is not your location but whether an EU reader sees the content.
The responsible entity is the “deployer”: an organization that employs AI tools and publishes content under its own name. The developer of the AI tool isn’t responsible. If you draft with an AI assistant and your team runs it, you’re the deployer.
“Public interest” is a broad term. It covers politics, public administration, justice, public health, the environment, and consumer safety, along with financial, scientific, or cultural developments open to public debate. Most marketing content sits outside that scope, but a significant amount of health, finance, and safety-adjacent content doesn’t.
As mentioned, failing to comply results in a maximum penalty of €15 million or three percent of worldwide annual turnover, whichever is greater. That ceiling only applies once a national regulator investigates and acts; it isn’t automatic.
National market surveillance authorities across the 27 member states share enforcement with the EU’s AI Office, which means the odds of scrutiny will vary by country and by how visible the content is. When you must disclose AI-generated content, the European Commission provides standard icons for labeling so the notice is consistent across the market, whether that’s a small tag on a webpage or a line in a video description.
Here’s the part most coverage of the new AI transparency rules skips. The disclosure rule doesn’t apply where the AI-generated text has been through a process of human review or editorial control and a named person or organization holds editorial responsibility for publication.
The problem is what counts as a review. The European Commission’s guidance is explicit: a review must be substantive, which means a deliberate examination of the content’s substance by someone with relevant knowledge and professional judgment. “Superficial, solely formal, or procedural checks” don’t qualify. A spell-check pass on its own doesn’t meet the standard, according to the Commission’s examples. Neither, in our view, would a second AI tool humanizing a first AI tool’s draft. The Commission hasn’t stated that directly, but it follows the same logic: no human with relevant expertise has examined the content for meaning.
This is where most teams run into trouble. A quick skim for typos and tone isn’t editorial control under the law. The expectation is a real examination of substance by someone qualified to judge it, with the ability to act on what they find.
There are four key factors to consider when assessing whether your review process meets the Article 50 standard:
The exemption isn’t something you apply for in advance, and no regulator preapproves a workflow. It’s something you invoke if a market surveillance authority asks why a piece of AI-assisted content carries no disclosure label – and the record you can produce at that moment is what determines if your exemption is valid.
Someone with relevant subject knowledge must have read the piece for meaning and framing and performed fact-checking. Fact-checking accuracy is a minimum requirement here, and the Commission’s guidance names it explicitly.
A proofread that only catches spelling, grammar, and typos doesn’t count, no matter how thorough. Neither does a read by someone without relevant expertise in the subject matter: a marketing coordinator signing off a piece on drug interactions isn’t substantive review.
✘ Doesn't Qualify
✔︎ Qualifies
The reviewer must have the authority to send the piece back or change it substantially. Merely suggesting edits for someone else to review isn’t sufficient. A review step with no power to stop or change the content isn’t editorial control; it’s a suggestion box. If the writer, not the reviewer, has the final say on whether a suggestion makes it in, the person responsible for editorial decisions isn’t truly responsible. This is a common structural problem in content teams that build reviews into the workflow as a courtesy step rather than a required checkpoint.
A specific person or entity must be on record as responsible for the published piece. The record can’t point to “the team” in general. Editorial responsibility needs a name attached, not a department.
This doesn’t mean a single person must review every piece but rather that any given content has an identifiable individual or organization responsible for it if challenged.
You must be able to show the review happened and not rely on your word against a regulator’s doubt. A dated sign-off or a documented workflow step all count as evidence. A verbal assurance that “someone always checks it” does not.
Evidence is the key component that protects the other three parts of the test: you can have the right reviewer, with real authority, clearly named, and still fail if there’s no record of what they did.
Go through the four-part test against your actual workflow. For each part, mark where you stand: Covered, At Risk, or Exposed.
Most teams land in “At Risk” somewhere, usually on evidence. The review itself is often reasonable; the paper trail behind it is not. Before you build that record from scratch, it’s worth knowing what an outsourced editorial team could save you versus building the equivalent process in-house.
If you marked “Exposed” on more than one part, your current process most likely wouldn’t hold up if a regulator asked to review it. That doesn’t mean starting over. It usually means naming a reviewer who already has the expertise and giving them explicit authority to reject or change a piece. You should also log the decision in a place easy to locate later. The gap between “Exposed” and “Covered” is often a workflow change, not a headcount one.
Subscribe to Beyond the Margins and get your monthly fix of editorial strategy, workflow tips, and real-world examples from content leaders.
Δ
None of this is new. Teams that already ran a true editorial layer, with named reviewers who could reject weak copy and a record of what changed and why, built it for quality reasons long before Article 50 existed. The regulation didn’t invent the standard, but it did give a deadline to an issue that careful publishers were doing anyway.
That’s worth remembering when the label itself feels like the focus. A Stanford study found that telling readers a piece of policy content was AI-generated didn’t meaningfully change the persuasiveness of the writing. The label alone isn’t what protects a reader or a brand’s credibility. The review behind the content provides that protection, regardless of a disclosure tag.
Viewed this way, the four-part test isn’t a compliance hurdle added onto content operations. It’s a description of what a functioning editorial process already looks like: qualified reviewers, real authority to change or reject content, a named owner, and a record worth keeping. Teams that treat the test as a foundation for good editorial practice tend to find that the gap between their current workflow and the test is smaller than expected.
The teams most exposed right now are ones that scaled AI-assisted content output faster than they scaled editorial oversight. A drafting process can grow tenfold with a new tool. A review process built around one or two people skimming for tone usually can’t grow comparatively, and it starts to show under the kind of scrutiny Article 50 invites. The fix isn’t slowing down production; it’s incorporating a review layer that can carry the current volume and produce a record when asked.
If you’re not sure whether your current process would hold up, that’s what our editorial process audit can determine. We’ll work through the four-part test with you and identify the gaps. Our analysis will help you close any gaps – whether that means changes to your internal workflow or the addition of a managed editorial layer that handles it for you. Book your free editorial process audit, and we’ll take it from there.
This article is operational guidance based on publicly available sources, not legal advice. Consult qualified counsel for advice specific to your organization’s circumstances.
Yes. The obligation applies if the content reaches EU readers, regardless of where the company producing it is based.
The exemption only matters where the disclosure obligation applies in the first place. Most commercial marketing copy falls outside Article 50(4) entirely – the obligation is for content published to inform people on a matter of public interest, not by AI use alone.
Public interest content is politics, public administration, justice, fundamental rights, public security, health, the environment, and consumer safety, among other topics that inform public decision-making.
No. The law requires substantive review, not a rewrite. The reviewer must have the authority to change or reject the piece when necessary.
Penalties may reach up to €15 million or three percent of global annual turnover, whichever is higher, once a regulator acts.
No. The guidance explicitly excludes superficial, solely formal, or procedural checks, including spell-checking, from counting as review.
No, but the evidence that supports it does. You cannot register for the exemption ahead of time. It’s something you rely on if a regulator ever challenges a piece of unlabeled content. A record of each review must exist from the moment of publication. You can’t reconstruct the record after the fact.
The deployer is whoever publishes the content under their own name and controls the AI system in that process. If you commission content from an agency but publish it on your own channels, you are most likely the deployer. The disclosure obligation, and the exemption, sits with you.
Using AI to generate content?
Let’s talk about the support you need.
We use cookies to give you the best possible experience with Proofed. Some are essential for this site to function; others help us understand how you use the site, so we can improve it. We may also use cookies for targeting purposes.