Since August 2, 2026, the transparency rules of Article 50 of the EU AI Act apply—and one question lands in my inbox weekly, usually in German: „Muss ich eine KI-übersetzte Übersetzung nach dem EU AI Act als KI-generiert kennzeichnen?“ The answer is binary, and it turns on the depth of your post-editing.
Short answer: is an AI-translated text subject to labelling?
No—reviewed AI translation is exempt. The Commission’s final guidelines of July 20, 2026 treat AI-generated translation of text as an assistive “standard processing” function under Art. 50(2), and the Art. 50(4) disclosure duty does not bite where the content has undergone human review or editorial control with a natural or legal person holding editorial responsibility. Substantive post-editing plus a named responsible editor means no label.
Yes—raw or lightly touched MT output must be disclosed if it is published to inform the public on matters of public interest. A formal click-through approval with no substantive review does not count as human review under the guidelines. From that point on, the text is synthetic content subject to marking.
One scope limit matters before anything else: Art. 50(4) covers text published with the purpose of informing the public on matters of public interest. It does not reach every internal memo that passed through an MT engine. Annual reports, sustainability statements, press releases, and investor-facing websites sit squarely inside that scope—which is where most of my DE→EN work lives.
Article 50: what has actually applied since August 2, 2026—and the myths that never existed
Article 50 is the transparency chapter of Regulation (EU) 2024/1689, and it switched on fully for transparency purposes on August 2, 2026. It imposes two different duties on two different parties, and most of the confusion I see in project teams comes from blending them:
- Art. 50(2)—a provider duty: makers of AI systems that generate synthetic text, images, audio, or video must mark the outputs in a machine-readable format so they are detectable as artificially generated. This duty does not apply to the extent the system performs an assistive function for standard editing or does not substantially alter the input data or its semantics. Machine translation of a human-written source text sits exactly in that carve-out, as the final guidelines confirm.
- Art. 50(4)—a deployer duty: whoever publishes AI-generated text to inform the public on matters of public interest must disclose that it is artificially generated—unless the content went through human review or editorial control with a named person holding editorial responsibility.
The myths are worth naming because they circulate freely in translation-industry LinkedIn posts. There is no general obligation to slap a visible “AI-generated” banner on every machine-translated sentence. There is no duty that applies to unpublished or purely internal translations. And there is no rule that hiring an LSP removes the deployer’s duty—the exemption comes from what was actually done to the text, not from who touched it.
The stakes are real but should not be dramatized out of proportion. Non-compliance with the Article 50 transparency obligations can trigger fines of up to €15 million or 3% of worldwide annual turnover. That number belongs in your compliance conversation; it is not a reason to abandon MTPE, which the same legal framework treats as the very thing that lifts the labelling duty.
The Commission guidelines of July 20, 2026: why AI translation counts as standard processing
The European Commission published its final guidelines on the Article 50 transparency obligations on July 20, 2026—a guidelines document plus a Commission communication approving the content. For translation buyers, two lists in those guidelines do the heavy lifting.
First, the guidelines list “AI-generated translations of text” among the standard-editing examples exempt from the machine-readable marking duty under Art. 50(2). Second, they list the “AI-supported translation of a human-written article whereby the translation has undergone human review” as a worked example that qualifies for the Art. 50(4) human-review exception. In other words: the two duties that could have hit machine translation both point the same way, as long as a human review stands behind the output. Per the Commission guidelines as analyzed by Locize, this was a deliberate clarification of the consultation draft from May—translation landed firmly on the exempt side.
One honest caveat, because the German-language commentary on the guidelines stresses it too: the guidelines are non-binding, and classification always depends on the function the system actually performed in the concrete case. An MT engine that substantially rewrites, summarizes, or generates content rather than translating it can fall outside the standard-processing idea. The exemption follows the work done, not the product category on the invoice.
What “substantive human review” means in post-editing practice
The guidelines define human review as “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.” That definition reads as if it was written with specialist translation in mind—and it sets a bar that not everything marketed as MTPE clears.
The guidelines are equally explicit about what fails the test: “Superficial, solely formal or procedural checks (e.g. spell-checking or grammatical correction) … cannot fulfil the conditions for human review or editorial control.” A light post-edit that fixes articles and prepositions but never asks whether the translated sentence is actually true to the source’s substance is, legally, the same as no review.
In my own projects, the distinction shows up as a question I can answer for every job: did someone with subject-matter competence open the source and the target side by side and take responsibility for the content—not just for the grammar? For an IFRS note disclosure, that means the reviewer understands what the consolidated statement says. For a CSRD report, it means the reviewer can tell a substantive claim from a boilerplate sentence. Full post-editing depth plus a named editor: exempt. Anything less: disclose or don’t publish.
This is also why industry voices position an MTPE process certified to ISO 18587 as complying with Article 50 without content labelling, while raw machine translation published without human review counts as synthetic content subject to marking. The certification is not a magic shield—the substance of the review is what carries the exemption—but a defined post-editing process with competent reviewers is exactly what the guidelines describe.
The evidence path that survives an audit
An exemption you cannot evidence is an exemption you do not have. Nothing in Article 50 prescribes a documentation format, which means standard professional hygiene does the job. If a regulator or an auditor ever asks “who reviewed this?”, the answer should be retrievable in minutes, not reconstructed from email threads. What I would want on file for every published, post-edited deliverable:
- Reviewer identity and qualification—who reviewed, with what subject-matter competence; not just “the language team.”
- Process description—which MT engine produced the draft, and that post-editing was performed to full review depth (an ISO 18587-style process description fits here).
- Timestamps—source received, draft generated, review completed, delivery sent. A plausible review window matters.
- Named editorial responsibility—the natural person at the client or publisher who signed off the publication. This is the Art. 50(4) linchpin.
- Sign-off statement—a short sentence in the delivery documentation: content substantively reviewed, named editor responsible.
None of this requires new tooling. A CAT-tool project history, a delivery note with reviewer initials, and a client-side approval email together carry you a long way. The only genuinely new habit for most teams is the last item: putting the sign-off in writing, on every job, before publication.
Where MTPE tips over into labelling territory—and liability
The exemption is not a property of the workflow label on the purchase order. It is a property of the delivered text. Three patterns I would flag immediately in any project review:
- Raw MT output in a published report. Someone pastes engine output straight into the annual report or the website and ships it. No review, no named editor, published to inform the public—disclosure duty triggered, and the publication is factually at risk regardless.
- Light post-edit without subject review. A reviewer polishes grammar and style but has no domain knowledge and never checks the translation against the source’s substance. Under the guidelines this is formally a procedural check—legally equivalent to no review.
- Unchecked fuzzy and TM matches. A translation-memory match from an earlier document is recycled untouched into a new report. If nobody verifies that the reused sentence still says something true in the new context, that segment has had no human review—whatever the CAT-tool match rate says.
That third case is the uncomfortable one for the industry, because it predates machine translation entirely. But the logic is unforgiving: the Art. 50(4) exception asks whether the published content underwent substantive human review, and a recycled match inherits no such review from a job done years ago. Review your matches like everything else.
What project managers and language leads should document now
If you run translation workflows on either side of the vendor relationship, the following SOP keeps you on the exempt side of the line without heroic effort. I would treat it as a one-page standard operating procedure, agreed with the client in the kickoff, not as a legal memo:
- Define your review depth per document class: full post-editing with source-target comparison for anything published or public-facing; document what lighter touch is acceptable for internal use.
- Name the reviewers and their subject-matter qualification in the project setup; keep the list current.
- Record MT usage per job: which engine, which segments, what the human step covered.
- Require a written sign-off before publication, with a named editorially responsible person on the client side.
- State the handling in the offer and order confirmation: “Machine translation with substantive human post-editing; reviewed content is exempt from AI-content labelling under Art. 50 EU AI Act.”
- Escalate raw-MT requests explicitly: if a client asks you to publish unreviewed engine output in scope of Art. 50(4), put the disclosure duty in writing before delivery.
That sixth point is the conversation guide in miniature. Most clients asking for “just run it through the machine” have not thought about the labelling question at all; a two-sentence explanation in the kickoff usually resolves it, and it positions the post-edited deliverable as the compliant option. If you want a template for that conversation, my services and pricing page describes how I scope MTPE, and the MTPE rate starts at €0.51 per standard line—60% of the €0.85 no-match rate, for a deliverable that carries the exemption with it.
The transition window until December 2, 2026: what MT engine and CAT-plugin users must check
One transitional rule is easy to miss. Providers of in-scope systems placed on the EEA market before August 2, 2026 have until December 2, 2026 to comply with the machine-readable marking duty for their pre-existing systems. Providers of newer systems—and all deployers—had no such grace period.
For translation teams, the practical to-do list before December 2, 2026 is short:
- Ask your MT engine vendor whether machine-readable marking of outputs is in place, and since when.
- Check the same for your CAT-tool MT plugins and any AI writing assistant wired into the workflow—each is potentially a provider in scope.
- Understand that this is a provider duty: your own review-and-editor setup remains the deployer-side story, and it has been live since August 2, 2026, with no deferral.
FAQ: the five most common questions on AI labelling and translation
Does the EU AI Act require a label on every machine-translated text?
No. The Art. 50(4) disclosure duty covers AI-generated text published to inform the public on matters of public interest, and reviewed translations are exempt even there. Internal documents, working files, and unpublished drafts are outside the scope entirely. The Art. 50(2) machine-readable marking duty is a provider duty and explicitly exempts AI-generated translations of text as assistive standard processing.
Does post-editing count as human review?
Only if it is substantive. The guidelines require deliberate examination of the content’s substance by people with relevant knowledge and professional judgement. Spell-checking or grammatical correction alone does not qualify, and a light post-edit without subject-matter review fails the same test. Full post-editing with source-target comparison and a competent reviewer qualifies.
Who has to disclose—the translator, the agency, or the publisher?
The disclosure duty under Art. 50(4) sits with the deployer—the party publishing the text to the public. As a translator or LSP you are a service provider in the chain; what you control is whether the deliverable qualifies for the human-review exemption. The named person holding editorial responsibility for the publication is the publisher’s side of the bargain.
Do I have to use the official EU icons?
No. The Commission’s EU icons for labelling AI-generated content are optional, and the Commission states plainly that their use alone does not establish legal compliance. Where disclosure is required, what matters is that the disclosure is clear and effective for the reader; the icons are one way, not the way.
What changes on December 2, 2026?
That date ends the transition period for providers of generative AI systems that were already on the EEA market before August 2, 2026: from then, their machine-readable marking duty applies to those pre-existing systems too. It does not postpone any deployer duty—those have applied since August 2, 2026.
If your CMS or site build supports structured data, the FAQ block above can be published with FAQPage schema so the individual answers are extractable:
<script type="application/ld+json">
{
"@context": "https://schema.org",
"@type": "FAQPage",
"mainEntity": [
{
"@type": "Question",
"name": "Does the EU AI Act require a label on every machine-translated text?",
"acceptedAnswer": {
"@type": "Answer",
"text": "No. The disclosure duty under Art. 50(4) covers AI-generated text published to inform the public on matters of public interest, and reviewed translations are exempt even there. The machine-readable marking duty under Art. 50(2) is a provider duty and exempts AI-generated translations of text as assistive standard processing."
}
},
{
"@type": "Question",
"name": "Does post-editing count as human review under the EU AI Act?",
"acceptedAnswer": {
"@type": "Answer",
"text": "Only if it is substantive. The Commission guidelines require deliberate examination of the content's substance by persons with relevant knowledge and professional judgement. Spell-checking, grammatical correction, or a light post-edit without subject-matter review does not qualify."
}
},
{
"@type": "Question",
"name": "Who has to disclose AI translation: the translator, the agency, or the publisher?",
"acceptedAnswer": {
"@type": "Answer",
"text": "The disclosure duty sits with the deployer, i.e. the party publishing the text to the public. A named person holding editorial responsibility for the publication is required for the human-review exemption."
}
},
{
"@type": "Question",
"name": "Do I have to use the official EU icons to label AI-generated content?",
"acceptedAnswer": {
"@type": "Answer",
"text": "No. The EU icons are optional, and their use alone does not establish legal compliance. What matters is a clear and effective disclosure where one is required."
}
},
{
"@type": "Question",
"name": "What changes on December 2, 2026 under the EU AI Act?",
"acceptedAnswer": {
"@type": "Answer",
"text": "December 2, 2026 ends the transition period for providers of generative AI systems placed on the EEA market before August 2, 2026. Their machine-readable marking duty applies to pre-existing systems from that date. Deployer duties were not deferred and have applied since August 2, 2026."
}
}
]
}
</script>Conclusion: honest MTPE is now a legal question—and a sales argument
The Article 50 regime draws a line that happens to run exactly where professional standards always ran it: between output a competent person stands behind and output nobody checked. Since August 2, 2026, that line is no longer just a quality argument. Reviewed translation is exempt from AI-content labelling; unreviewed output published to the public is not, and can cost up to €15 million or 3% of worldwide turnover.
For LSPs and in-house language teams, this is the rare compliance story that plays in your favor. The work you were already doing—substantive post-editing by people who understand the subject matter, with a named sign-off—now has a legal payoff your clients can put in their audit file. Document it, name it, and say it in the kickoff. If you want to review how a compliant MTPE workflow looks in a concrete DE↔EN project, you can reach me via the contact page.