Is machine-translated content "AI-generated" under the EU AI Act?

From August 2, 2026, Article 50 of the EU AI Act requires AI-generated content to be identifiable as such. Whether a machine translation of human-written text counts as "AI-generated" is one of the questions the Act never answers explicitly, and for anyone running a localized website or app it is the question that decides whether these rules apply to them at all.

Update, July 2026: when we first published this article, the Commission's draft guidance pointed to yes, and we flagged that the guidelines were draft and might shift. They did. The final guidelines of July 20, 2026 (C(2026) 5054) moved AI translations to the exempt side of the marking obligation, and made a human-reviewed AI translation their worked example of the Article 50(4) exemption. This article has been updated throughout; the practical to-do list at the end survived the reversal unchanged, which was the point of building it on the exemption rather than on the unsettled question.

This is not legal advice. This article summarizes public legal texts and guidance for orientation. The guidelines cited here are final since July 20, 2026 but non-binding; authoritative interpretation of the AI Act rests with the CJEU. For decisions about your own obligations, consult a qualified lawyer.

---

What Article 50 actually requires

Article 50 of the AI Act (Regulation (EU) 2024/1689) contains two separate obligations that matter for content, and they fall on different actors:

1. Machine-readable marking, on the AI system provider. Article 50(2) requires that providers of AI systems generating synthetic audio, image, video or text content

"ensure that the outputs of the AI system are marked in a machine-readable format and detectable as artificially generated or manipulated."

This is watermarking and metadata, not a visible banner. It is the obligation of whoever provides the AI system (the MT engine, the LLM), and it explicitly does not apply

"to the extent the AI systems perform an assistive function for standard editing or do not substantially alter the input data provided by the deployer or the semantics thereof."

Keep that exception in mind. It is where the whole machine-translation question lives.

2. Visible disclosure, on the deployer. Article 50(4), for text, requires that deployers of an AI system

"that generates or manipulates text which is published with the purpose of informing the public on matters of public interest shall disclose that the text has been artificially generated or manipulated."

This one is a visible label, but its scope is narrow (more on that below), and it has the exemption this article is ultimately about.

The dates, including the Omnibus trap

If you saw May 2026 headlines saying "the EU delayed the AI Act": the delay is real, but it is about high-risk systems. The part relevant to content and translation teams was not delayed.

The Omnibus changes are a provisional political agreement (May 2026) still awaiting formal adoption, as law firm briefings like Gibson Dunn's summary note. The August 2, 2026 date for Article 50 is not part of those changes: it stands in the Act as adopted. Non-compliance can be fined with up to EUR 15 million or 3% of total worldwide annual turnover, whichever is higher.

---

Is machine translation "AI-generated"? Asked and, since July 20, answered

Here is the genuinely interesting legal question. Machine translation does not create content from nothing: a human wrote the source text, the AI only moves it into another language. Is the result "artificially generated or manipulated" content?

The case for "no"

The strongest argument that machine translation falls outside Article 50(2) is the Act's own exception, quoted above: systems that "do not substantially alter the input data provided by the deployer or the semantics thereof" are exempt from marking. A faithful translation is, by definition, meaning-preserving. If the semantics survive the language transfer intact, the argument goes, the output is the deployer's human-authored content in a different surface form, not synthetic content. Translation tools also long predate generative AI, and nothing in the Act's recitals singles them out.

The case for "yes"

The counter-argument: every sentence an MT system outputs is text the model generated. Word choice, register, sentence structure, terminology: all of it is model output, and all of it can be wrong in ways that change meaning. A translation is not an edit of the source text; it is new text in a language the source never existed in.

This was the reading the Commission's draft guidelines of May 8, 2026 took: they listed "AI-generated translations and summaries of text" as their very first example of semantic changes requiring marking.

The final guidelines picked "no" for translations

The final guidelines of July 20, 2026 resolve the question the other way. The consultation pushback on exactly this point (documented in briefings like Covington's 10 takeaways) landed: in the final examples of "standard editing and minor alterations benefitting from the exception under Article 50(2) AI Act", the Commission now lists

"Grammar correction and spellchecking, linguistic and minor stylistic polishing that do not change the substance, meaning, style or messaging of text, AI-generated translations of text"

while the marking-required list keeps

"AI-generated summaries of text; paraphrasing or rewriting text that changes style, structure and meaning beyond mere grammatical and minor stylistic correction"

So a faithful AI translation is treated as standard editing of human-authored content: no machine-readable marking duty. A summary or a free rewrite is not. The boundary is the general test the guidelines state for the exception: the output must not change the content "in a material way" affecting its meaning, style or intent, assessed case by case. A translation that edits, condenses, or re-angles the source on the way through can cross back into marking territory.

Where that leaves you

Three caveats keep this from being a free pass:

• The guidelines are final but non-binding. They are the Commission's stated reading, the reference point for market-surveillance authorities; the last word belongs to the CJEU.
• The marking duty in Article 50(2) sits with the AI system's provider anyway, not with the company using MT for its website. As a deployer your direct exposure is Article 50(4). The provider side is organizing accordingly: the initial signatory list of the Commission's transparency Code of Practice (July 31, 2026) counts 83 providers, including OpenAI, Google, Mistral, Anthropic, Microsoft and Meta.
• And Article 50(4) is narrow, but alive for translation. The disclosure duty only covers text "published with the purpose of informing the public on matters of public interest": think public health information, safety guidance, reporting on political or economic developments, investor information. The guidelines explicitly exclude fiction, entertainment, and ordinary advertising (unless it makes health, safety or sustainability claims). Most localized UI strings and product copy never enter its scope, but an AI-translated help-center article on product safety, a translated corporate report, or translated public-service information can.

A useful summary of the obligations is the Article 50 text and the practical guide on artificialintelligenceact.eu. For the content that does fall under Article 50(4), everything turns on the exemption.

---

The exemption that matters: human review and editorial responsibility

Article 50(4) does not apply, verbatim,

"where the AI-generated content has undergone a process of human review or editorial control and where a natural or legal person holds editorial responsibility for the publication of the content."

The final guidelines construe this narrowly, and the wording is worth reading closely. Human review means

"the deliberate examination of the substance of the content by one or more natural persons possessing relevant knowledge and professional judgement pertaining to the subject matter under scrutiny"

with fact-checking the accuracy of the content named as "a minimum requirement that should be part of that review". And explicitly not:

"Superficial, solely formal or procedural checks (e.g. spell-checking or grammatical correction), the mere existence of an editorial policy, automated review processes or cursory editorial approval without substantive engagement by the human reviewer or the editorial entity, cannot fulfil the conditions for human review or editorial control."

Two further points are easy to miss. First, order matters: if AI modifies, supplements or reformulates the content after editorial sign-off, the exception becomes void; review has to be the last substantive touch before publication. Second, the second condition, editorial responsibility, means a named natural or legal person holds ultimate responsibility for the publication, and the guidelines expect their identity and contact details to be publicly findable (a website's terms or legal notice is given as an example).

And translation is not a hypothetical here: the guidelines' own list of examples that satisfy the exception includes, verbatim,

"AI-supported translation of a human-written article whereby the translation has undergone human review."

That sentence is this article's whole subject, resolved in one example.

What the guidelines effectively demand is not a feeling of having reviewed things, but a documented editorial workflow with identified responsible persons. If a regulator asks, "this went through human review" needs to be demonstrable: who reviewed it, what they could change, who approved publication.

For deployers who want a formal vehicle on top, the Commission's Code of Practice on transparency of AI-generated content was assessed as adequate under Article 50(7) precisely for demonstrating these conditions, and adherence is open to deployers (Section 2 of the Code). As of September 9, 2026, the signatory list names the large model and platform providers whose engines translation workflows commonly use, and one translation-technology company under Section 1, but no other localization vendor by name; consistent with faithful translations sitting outside the marking duty, and a reminder that adherence is voluntary: staying off the list implies nothing about compliance. A documented review workflow is the evidence layer either way.

---

How to be safe under either reading

This is the practical part, and it is shorter than the legal part, because it was built to survive either answer to the MT question, and it did:

1. Route machine and AI translations through human review. Not necessarily all of them; review where it matters and where confidence is low (see below).
2. Make the review substantive. Reviewers need competence in the language and the subject matter, and the authority to approve, change, or reject the translation. A rubber-stamp click does not meet the bar the draft guidelines set.
3. Record who reviewed and approved what, and when. The exemption is only worth something if you can show the process happened.
4. Name who holds editorial responsibility for published content, and make that publicly findable (legal notice, terms, colophon).
5. Ask your MT/AI providers about machine-readable marking. Under Article 50(2) that part is their obligation for generative output; per the final guidelines, faithful translation output is exempt from marking, while summaries and rewrites are not.

The logic was deliberately either-way, and the reversal proved it out:

• For translated text that falls under Article 50(4)'s public-interest scope, your reviewed content sits squarely in the human-review exemption, which the final guidelines illustrate with exactly this case.
• For everything else, you have lost nothing: the same workflow is simply how you catch the mistranslations that were going to embarrass you anyway.

Compliance pressure and translation quality point at the same workflow. That is rare, and worth using.

---

Where Locize fits

Locize is built around exactly the workflow the exemption path needs:

• The review workflow turns translations into pending proposals that a human reviewer with the right permissions must accept or reject: deliberate, substantive engagement, per segment.
• Quality Estimation scores every AI translation with a confidence estimate, and the Review AI workflow routes low-confidence translations into review automatically. Your reviewers' substantive attention goes where the AI is least sure, instead of being spread thin across everything. (Background in the Quality Estimation announcement.)
• The history of every segment records when and how each translation changed and who changed or approved it, so "this went through human review" is something you can show, not just assert. The provenance export packages this evidence in one action, per version, language or namespace: a report as CSV and JSON listing who reviewed or approved each translation and when, the AI confidence score at the moment of the decision, what is still pending review, and a summary header so the file stands on its own in front of an auditor.
• User management lets you give reviewers and approvers distinct roles, which maps cleanly onto "identified responsible persons".

To be clear, and in the spirit of the disclaimer above: no tool makes you compliant with the AI Act, and Locize does not either. What it gives you is the documented, substantive, person-attributable review process that the Article 50(4) exemption asks for, the process the final guidelines illustrate with a human-reviewed AI translation.

If you want your translation workflow on that footing: create a free Locize account and enable the review workflow and Quality Estimation for your languages. The step-by-step setup, including the provenance export and what each review status means for an auditor, is documented in Running an Article 50(4)-compatible review workflow in Locize.

The question in the title has an official answer now. Whether your workflow is on the right side of it is the part you control.