As of 2 August 2026, Article 50 of the EU AI Act is fully in force. Companies publishing multilingual content, from e-commerce catalogues to financial reports, must now comply with new . On 20 July 2026, the European Commission published its final guidelines on Article 50, together with the Code of Practice on Transparency of AI-Generated Content, signed by around 190 organisations by the end of July. For localisation, marketing and legal teams, the question is a practical one: must a text translated using artificial intelligence always carry a visible label?
According to the Commission’s own guidelines, the answer is often no. But only if the workflow meets specific conditions. Let us look at what the EU AI Act and watermarking of multilingual content mean in practice, what the rules say about machine translation and synthetic dubbing, and why human review remains the safest approach.
What Article 50 really says about translating and localising content
Article 50 of EU Regulation 2024/1689 distinguishes between two parties with different obligations, and it is important not to confuse them. The provider of a generative AI system must mark its outputs, whether text, audio, video or images, in a machine-readable format that makes them detectable as AI-generated (technical watermarking). This obligation, laid down in Article 50(2), applies to the developer of the machine translation or synthetic dubbing technology, not to the company using it to localise a product catalogue or corporate video.
The deployer, meaning the company publishing the translated content, has different, more targeted obligations. Article 50(4) requires disclosure when content is a deepfake, or when AI-generated text is published to inform the public about matters of public interest, such as health, safety or investor information.
A watermark is not an overlaid caption but an invisible signal embedded in the content itself: in the pixels of an image, the waveform of an audio recording or the statistical structure of a text. It is generally accompanied by digitally signed metadata and allows software and platforms to recognise an output as AI-generated, even when nothing is visible to the naked eye. This is precisely the machine-readable marking that the EU AI Act, applicable from August 2026, requires providers of generative AI systems to implement.
For multilingual localisation, this distinction is crucial: technical watermarking is the responsibility of the technology provider, not the in-house marketing or legal department. The guidelines also clarify that systems already on the market before 2 August 2026 have until 2 December 2026 to implement machine-readable marking.
When an AI translation requires disclosure, and when it does not
This is the most useful point for localisation professionals. According to the European Commission’s guidelines, a faithful translation of a human-written text that has undergone human review or editorial control and is published under the editorial responsibility of a natural or legal person may qualify for the exemption from the disclosure requirement under Article 50(4). Disclosure means an explicit statement, visible to the public, that content has been generated or modified using artificial intelligence, for example through a label or information notice. Unlike an invisible watermark, disclosure is intended directly for the reader, listener or viewer. This is the explicit example cited by the Commission in its final guidelines of 20 July 2026, which classify faithful translation as standard editing rather than a substantial alteration requiring marking.
However, human review should not be treated as a mere formality. The content must genuinely undergo review or editorial control, and editorial responsibility for publication must be attributable to a natural or legal person. Entrusting this control to a professional with appropriate linguistic and subject-matter expertise is also good practice for ensuring content quality and reliability.
This means that not all machine translations are treated equally under the law. A raw translation published without checks may fall within AI transparency obligations, particularly if it concerns matters of public interest. A translation that has undergone review or editorial control, with clear editorial responsibility for publication, may instead qualify for the exemption from disclosure under Article 50(4), provided the relevant conditions are met.
Human-in-the-Loop: how a well-designed post-editing workflow protects your company
In translation, the Human-in-the-Loop approach is commonly known as machine translation post-editing, or MTPE. It is not a bureaucratic formality to be added at the end of the process. It is a structured workflow combining the speed of multilingual machine translation with the professional judgement of a qualified human reviewer before publication.
To stand up to potential scrutiny, a post-editing workflow should include several key elements. First, a language reviewer with expertise in the target language and the subject matter, rather than a generic check. Second, clear traceability of who reviewed what and when, with versions retained. Third, editorial responsibility explicitly assigned to an identifiable person or company, never vaguely attributed to the software or technology provider.
A linguistic project manager coordinating this process is more than an operational support role. They ensure that translated content meets the standards required by the Commission’s guidelines, from selecting the reviewer to performing the final checks. Documenting every stage of quality control therefore also helps protect the company, as well as being good editorial practice, including in the event of a challenge by a supervisory authority.
What are the risks of non-compliance?
Penalties for non-compliance with Article 50 are far from symbolic. They can reach EUR 15 million or 3% of annual worldwide turnover, whichever is higher. For an internationally active company publishing content in many languages and markets simultaneously, this represents a real financial risk, not merely a theoretical or remote one.
There is also a less visible but equally costly long-term risk: the trust of multilingual audiences. Content translated without checks, perceived as automatically generated and published without adequate transparency, can damage a brand’s reputation in overseas markets. In sectors such as healthcare, finance or economic reporting, this kind of damage can quickly lead to lost customers and diminished credibility among local partners.
It is always advisable to check with your legal department which specific obligations apply to your circumstances, as each type of content, market and target audience may present nuances that differ from the general rule.
How we help companies manage AI-generated multilingual content in compliance with the rules
For years, Aglatech14 has worked with the localisation, marketing and legal teams of Italian companies operating internationally on large-scale multilingual projects. Our approach combines high-quality machine translation with structured human review, supported by dedicated linguistic project managers and quality-control processes documented at every stage.
We design post-editing workflows tailored to every type of content: e-commerce websites, financial reports, technical documentation and marketing materials intended for multiple markets. Every stage remains traceable, with clearly assigned editorial responsibility and reviewers identified by language and subject area.
This does not replace specific legal advice, and every company should assess its own exposure with its legal department. Nevertheless, a well-designed human review process remains, in most cases, the simplest way to manage AI-generated multilingual content without intrusive disclaimers or unnecessary risks.
Official resources:
- Transparency obligations under Article 50 of the AI Act, European Commission
- Code of Practice on Transparency of AI-generated Content
Aglatech14 ContenTalkers
