AI Act Article 50: What Breaks the Editorial Exception After Sign-Off
An editor reviews a draft, checks the sources, and signs off. That should settle whether the text needs an AI-disclosure label under the EU AI Act.
It doesn't, by itself. Between sign-off and publication, the same text often passes through a translation step, an SEO pass, a summary for a newsletter, or a personalization layer — and each of those can either leave the earlier review intact or reopen the question the review was supposed to close.
Article 50(4) of the AI Act protects a specific object: the text a named person actually reviewed and takes editorial responsibility for. It does not protect every later version of that text by association.
In short
Article 50(4) requires deployers to disclose AI-generated or manipulated text published to inform the public on matters of public interest — unless the text underwent substantive human review and a natural or legal person holds editorial responsibility for the publication.
The European Commission's final guidelines, published 20 July 2026, read that exception narrowly. A spelling pass or a quick approval does not qualify. Fact-checking the substance is treated as a minimum requirement of a genuine review.
The guidelines also say that a substantive AI intervention after the review makes the exception unavailable. They do not list which post-approval steps count as substantive for Article 50(4), but they do classify AI steps for the Article 50(2) marking duty: faithful translation sits with grammar correction as standard editing, while summaries and paraphrases that change meaning require marking. Using that line as a working test for Article 50(4) is a reasonable reading, not a ruling. A substantive automated intervention after human approval requires a renewed review of whether the editorial exception still applies to the version that actually gets published.
Two obligations, two different actors
Article 50 splits the work by role. Article 50(2) sits with the provider — the organization that builds or supplies the generative AI system — and requires machine-readable marking of synthetic output, so a downstream system can detect it.
Article 50(4) sits with the deployer — whoever uses that system under its own authority and publishes the result. A publishing house can be both: a deployer when it uses a third-party writing tool, and a provider if it commissions and ships a system under its own brand.
The two obligations do not share a deadline. Article 50(4) has applied since 2 August 2026, on schedule. Only the Article 50(2) marking requirement got relief: Regulation (EU) 2026/1744, the Digital Omnibus on AI, added a transitional rule giving generative systems already on the market before 2 August 2026 until 2 December 2026 to implement machine-readable marking. The same regulation pushed back the high-risk obligations — to 2 December 2027 for Annex III systems and 2 August 2028 for Annex I — but left the deployer duty in Article 50(4) and its application date unchanged.
What a review has to look like
The Commission's FAQ on Article 50 sets two cumulative conditions for the exception, and both have to be met by the same publication.
Human review means deliberate examination of the substance by someone with relevant competence and professional judgment — the guidelines point to peer review and professional validation chains as the model. Editorial responsibility means a specific natural or legal person holds ultimate legal responsibility for that publication, including for the review itself.
What does not count: spell-checking, a grammatical pass, or a cursory approval with no substantive engagement. A rubber-stamp click on a shared drive does not meet the bar the guidelines set, however fast the workflow makes it feel.
Standard editing keeps the exception. Substantial alteration reopens it.
This is the distinction publishers are most likely to get backwards. The final guidelines moved AI-generated translation out of the "requires marking" list and into the same category as grammar correction and minor stylistic polishing — provided the translation is faithful and does not change the substance, meaning or messaging of the text. Summaries and substantive rewrites stayed on the other side of the line.
Post-approval step | Category under Art. 50(2) guidance | Likely effect on the Art. 50(4) exception |
|---|---|---|
Grammar and spellcheck | Standard editing | Original review still covers the published text |
Faithful AI translation | Standard editing | Original review still covers the published text |
SEO pass that changes headings, structure or claims | Substantial alteration | Published version needs its own review |
AI-generated summary | Substantial alteration | Published version needs its own review |
Personalization that changes wording per reader segment | Substantial alteration (case by case) | Published version needs its own review |
The practical rule follows directly: any automated step that changes structure, meaning or claims after sign-off is a new publication event for Article 50(4) purposes, whatever happened to the draft before it.
The record worth keeping
None of this is enforceable from memory. For each publication that could plausibly be read as informing the public on a matter of public interest, a publisher needs to be able to produce, on request: the version a named reviewer actually approved, who that reviewer was, when the review happened, and what — if anything — an AI system did to the text afterward.
That last element is the one most workflows don't capture today, because the tools that perform the post-approval step — a CMS plugin, an SEO assistant, a translation pipeline — sit outside the editorial system that logged the review. Technical marking has the same weak point: metadata and Content Credentials attached at generation time can be stripped or corrupted on export to a CMS or on republication, which does not shift the deployer's Article 50(4) obligation but does remove the evidence a publisher would otherwise rely on. This is a general limitation of metadata-based provenance, worth testing against the specific CMS and export formats a given publishing house actually uses rather than assuming it away.
This is the same gap Citations Logic is built to close on the usage-evidence side: a record tied to the event, not to a tool's own reporting layer.
What non-compliance actually costs
Article 99 sets the numbers. A breach of Article 50's obligations falls under the same tier as most operator obligations: up to €15 million or 3% of total worldwide annual turnover, whichever is higher. Supplying incorrect, incomplete or misleading information to a notified body or a national competent authority in response to a request is a separate, lower tier — up to €7.5 million or 1% of turnover. SMEs and start-ups get the more favorable calculation: the lower of the two figures, not the higher.
Market surveillance authorities enforce both tiers from 2 August 2026. The Commission's guidelines are not binding law, but they are what those authorities are expected to apply first.
Frequently asked questions
Does translating an AI-drafted article after human review require a new disclosure check?
Not on its own, on the most reasonable reading. The final guidelines classify faithful AI translation as standard editing for the Article 50(2) marking duty; applied to Article 50(4), that suggests a faithful translation is not the substantive intervention that voids the editorial exception. A translation that also rewrites or restructures the content is.
Does an SEO optimization pass count as a substantial alteration?
It depends on what changes. An SEO pass limited to metadata is unlikely to reopen the question. One that rewrites headings, restructures sections or adjusts claims to target a query is a substantive change under the guidelines, and the published version should be reviewed again.
Who is responsible if a vendor's AI tool touches the text after a publisher's internal review?
The deployer — the organization that publishes the text — carries the Article 50(4) obligation regardless of which tool performed the intervention. Vendor contracts should specify what a tool is permitted to change after sign-off and require notice when it does more than that.
Is a disclosure statement the same as proof that the editorial exception applies?
No. A statement records what someone says happened. Demonstrating the exception requires being able to show which version was reviewed, by whom, when, and what happened to the text afterward — the same distinction between disclosure and audit trail that applies to AI use in STM publishing.
Continue the evidence chain
For the licensing side of the same regulation, see EU AI Act and content licensing.
For how a disclosure statement differs from proof of what happened, see AI disclosure and audit trails for STM publishers.
For why static transparency rules age out and what should replace them, see AI transparency law and durable evidence.
Book an AI usage evidence assessment
Sources
European Commission — FAQ on transparency obligations under Article 50 of the AI Act
Transparency obligations under Article 50 of the AI Act
European Commission — Final guidelines on transparency obligations for providers and deployers of AI systems (20 July 2026)
Guidelines on transparency obligations
Regulation (EU) 2026/1744 of 8 July 2026 (Digital Omnibus on AI)
Official Journal text
EU Artificial Intelligence Act — consolidated text of Article 50 and Article 99
Article 50 · Article 99
C2PA — Frequently Asked Questions
C2PA FAQ
This article is operational analysis for publishing and editorial teams, not legal advice. Confirm application to a specific publication with counsel.