To comply with EU AI Act Article 50, providers of AI translation systems must mark outputs in a machine-readable format that is effective, interoperable, robust, and reliable. As of May 2026, no single technique meets all four criteria alone — the European Commission's draft Code of Practice (published May 8, 2026, with final guidance expected in early June) recommends combining at least two approaches: cryptographic provenance metadata (the C2PA standard) plus a complementary signal such as a watermark or logging method. The rules become applicable on August 2, 2026.
This is a technical implementation guide for product, legal, and compliance teams at companies that build with or deploy AI translation. It is not legal advice — confirm specifics with counsel before shipping any compliance posture.
Bluente is an AI-powered document translation platform used by 30,000+ professionals to translate files in 120+ languages while preserving original formatting. The marking question matters to anyone shipping AI translation into the EU, whether as a vendor, an embedded API user, or a deployer of AI-translated public content.
What Does Article 50 Actually Require for AI Translations?
Article 50(2) of the EU AI Act requires that providers of AI systems generating synthetic text, audio, image, or video content ensure outputs are "marked in a machine-readable format and detectable as artificially generated or manipulated." Article 50(4) further requires deployers of AI text generation systems to disclose when text intended to inform the public on matters of public interest has been AI-generated or manipulated.
The European Commission's draft Article 50 guidelines, published May 8, 2026, explicitly classify AI translation as a "substantive change" — meaning AI-translated text falls within the marking obligation. Translations are not exempt because they are derivative; they're treated as new AI-generated output. The transparency rules apply from August 2, 2026.
Which Marking Techniques Does the Code of Practice Recognize?
The draft Code of Practice recognizes five categories of marking technique: watermarks (visible or imperceptible signals embedded directly in the output), metadata identifications (provenance information attached to the file), cryptographic methods (digital signatures and content credentials proving provenance), logging methods (audit-trail records linking output to a generation event), and fingerprints (statistical signatures that identify content even after re-encoding).
The Code states that as of mid-2026, no single technique simultaneously meets all four Article 50(2) criteria — effectiveness, interoperability, robustness, reliability. A compliant marking posture therefore combines techniques. The most commonly recommended combination for text and document output is C2PA-compliant provenance metadata plus a robust logging method maintained by the provider.
What Is C2PA and Why Does It Matter for Translations?
C2PA (Coalition for Content Provenance and Authenticity) is an open standard for attaching cryptographically-signed provenance metadata to digital content. The metadata records the content's origin, the AI system that generated it, the timestamp, and any edits. For documents, C2PA manifests can be embedded in PDF metadata, DOCX core properties, and image EXIF/XMP fields without changing the visible content.
C2PA is the most likely interoperable baseline for Article 50 compliance because it is already adopted across the major AI platforms (OpenAI, Microsoft, Adobe, Google) and the major content authoring tools. For AI translation output, embedding a C2PA manifest in the translated file — declaring "this content was machine-translated by [system], from [source language] to [target language], on [timestamp]" — meets the metadata identification criterion.
How Do You Implement This for AI Translation Output?
For a typical AI translation workflow, the implementation has three parts. First, embed a C2PA manifest in the output file at the moment of translation, signed with the provider's certificate, declaring that the translation is AI-generated and naming the source content. Second, maintain a provider-side log linking the output's hash to the generation event — this satisfies the logging method criterion and supports robustness against metadata stripping. Third, for public-facing content under Article 50(4), the deployer adds a visible disclosure (a footer, banner, or document property) that the translation was AI-generated.
For document formats specifically, C2PA manifests embed cleanly in PDF, DOCX, XLSX, PPTX, and image files. The original file's content and formatting remain untouched; only the metadata layer is updated. Format-preserving translation platforms can attach the manifest as part of the output pipeline.
What About Translations That Are Then Edited by Humans?
The draft Code of Practice treats human post-editing as a substantive intervention. If a human reviewer edits an AI translation before publication, the provenance chain can be updated to reflect the human edit, and in many cases the visible-disclosure obligation under Article 50(4) is relaxed when the final output reflects substantial human authorship. The exact threshold for "substantial" is not yet finalized, and counsel should weigh in for high-stakes public content.
For mixed AI-human workflows — common in legal, regulatory, and high-stakes business translation — the practical posture is to keep the C2PA manifest current at each stage: AI translation event, human review event, final approval event. The audit trail is what matters more than any single label.
Does This Apply to Internal Documents and B2B Workflows?
Article 50(4) — the visible-disclosure obligation — applies specifically to text intended to inform the public on matters of public interest. Internal corporate communications, private contracts, and B2B documents shared bilaterally are typically outside its scope. Article 50(2) — the machine-readable marking obligation — applies to the providers of the AI translation system at the point of generation, regardless of downstream use.
For most professional translation use cases — translating a contract for in-house counsel, a financial statement for a banking team, a regulatory submission for a compliance group — the practical implication is that the AI translation provider must mark the output. Whether the deployer (the firm using the translation) must add visible disclosure depends on what they do with it.
What About Voice and Real-Time Translation?
The Code of Practice extends the same logic to synthetic audio. Real-time AI translation — voice-to-voice or document-to-voice — falls within Article 50(2). The marking expectation is harder to implement live: cryptographic watermarks in audio streams are an active research area, and provenance metadata typically attaches to the recording rather than the stream. Most vendors are converging on logging-based compliance for live audio, with provenance attached when a recording is exported.
How Does Bluente Approach Article 50 Compliance?
Bluente is preparing for the August 2026 transparency rules by combining C2PA-compliant provenance metadata embedded in translated files with a provider-side audit log linking every output to its generation event. The architecture sits on the same compliance foundation as Bluente's existing certifications: SOC 2 Type II, GDPR, and ISO 27001. Zero data retention applies on the user-content side; the audit log records the generation event and metadata, not the document content.
For Bluente customers handling EU public-facing translations under Article 50(4), the platform makes it straightforward to attach a visible disclosure to the output file via document properties or a configurable footer line. For internal or B2B use, the embedded provenance metadata is sufficient on its own.
Frequently Asked Questions
Q: When does the marking obligation actually start? August 2, 2026, when the Article 50 transparency rules become applicable. The Code of Practice is expected to be finalized in early June 2026, two months ahead of the deadline.
Q: Does this apply outside the EU? The EU AI Act applies to AI systems placed on the EU market or whose output is used in the EU. Translation providers outside the EU whose service is used by EU customers fall within scope. Most major AI translation vendors are taking a global compliance posture rather than fragmenting by jurisdiction.
Q: Can I just add the word "Translated by AI" to a footer and call it done? A visible footer disclosure may satisfy Article 50(4) for public-facing content, but it does not satisfy Article 50(2)'s machine-readable marking requirement. Compliance requires both: a machine-readable signal (typically C2PA metadata) plus, for public-interest content, a visible disclosure.
Q: What's the enforcement risk if I don't mark translations? Article 50 enforcement penalties under the EU AI Act can reach up to €15 million or 3% of annual global turnover for non-compliance. Enforcement priorities for the first 12 months are expected to focus on high-impact public-facing AI content rather than internal B2B workflows, but the rule is the rule.
Q: Does C2PA add visible markers to my document? No. C2PA manifests are metadata only. The document's visible content and formatting are unchanged. The provenance is readable by tools that parse the manifest and invisible in normal document viewing.
Q: Where can I read the actual draft Code of Practice? The European Commission published the draft on May 8, 2026 via the AI Office. The consultation period runs through June 3, 2026, with final guidance expected before the August 2 application date.
Start translating documents for free. Bluente preserves your formatting across 120+ languages in under 2 minutes. Try BluTranslate free — no credit card required.
This article is technical and compliance guidance, not legal advice. Confirm Article 50 implementation choices with counsel before deploying in production.