EU AI Act: what changes does it bring for translation

On 2 August 2026, the EU AI Act takes its most significant step yet toward full enforcement. After months of staggered implementation, Article 50 — the transparency framework at the heart of the regulation — becomes fully applicable, along with the sanctions regime that gives it teeth. For any company that uses artificial intelligence to produce, translate or distribute content reaching audiences in the European Union, the rules have changed. This article explains what has changed, who is affected, and — with specific focus on translation and localisation — what it means in practice.

The EU AI Act: a phased regulation reaching its critical threshold

The AI Act (Regulation EU 2024/1689) entered into force in August 2024, but the European legislator designed its application in deliberately staged phases, calibrated to the level of risk each type of AI system carries. Understanding the timeline is essential to understanding why 2 August 2026 matters.

What was already in force before August 2026

• 2 February 2025 — Prohibited practices and AI literacy. Article 5 banned eight categories of AI use deemed to carry unacceptable risk: mass biometric surveillance in public spaces, social scoring by public authorities, subliminal manipulation, untargeted scraping of facial recognition databases, emotion recognition in workplaces and schools, and others. Simultaneously, Article 4 made AI literacy a legal obligation: deployers must ensure staff working with AI systems have a sufficient level of competence. This obligation has been in force for over eighteen months and remains widely unmet.
• 2 August 2025 — GPAI models and governance architecture. Rules governing General Purpose AI models — the large foundation models that power most generative AI tools — became applicable, alongside the European governance structure and the designation of national supervisory authorities. In Italy, the Agenzia per la Cybersicurezza Nazionale (ACN) was confirmed as the competent authority under Legge 132/2025.

What comes into force on 2 August 2026

The 2 August 2026 deadline activates three interconnected blocks of obligations that together constitute the main operational framework of the regulation for the vast majority of businesses.

• Article 50 — Transparency obligations. The requirement to disclose AI use to end users, to label synthetic content in machine-readable format, and to identify AI-generated text published in the public interest. This is the provision with the broadest reach: it applies not only to high-risk systems but to any organisation deploying AI in the four scenarios Article 50 covers.
• High-risk systems (Annex III). The full obligations for AI systems classified as high-risk — including conformity assessments, technical documentation, human oversight requirements and registration in the EU database — apply to systems in areas such as employment, education, access to public services, law enforcement and border management. Note: following the EU Digital Omnibus provisional agreement of May 2026, some high-risk obligation timelines have been partially revised.
• Full sanctions regime (Article 99). Administrative fines become enforceable: up to €35 million or 7% of global annual turnover for violations of prohibited practices, and up to €15 million or 3% of global annual turnover for infringements of other obligations, including those under Article 50. The lower tier — €7.5 million or 1% — applies to the provision of incorrect information to authorities.

What follows after August 2026

• 2 December 2026 — Transition deadline for providers of generative AI systems already placed on the market before 2 August 2026 to implement the machine-readable content marking requirements under Article 50(2).
• 2 August 2027 — High-risk systems embedded in products already regulated under existing EU directives (Annex I: machinery, medical devices, automotive, aviation, and others) must comply with the full AI Act framework.

Article 50 in detail: the four transparency obligations

Article 50 is the provision most companies will encounter directly, regardless of whether their AI use involves high-risk applications. It establishes four distinct obligations depending on the type of AI use involved.

1. Disclosure of AI interaction (chatbots and virtual assistants)

Any deployer of an AI system that interacts directly with people — a chatbot, a virtual assistant, an automated customer support agent — must ensure that users are clearly informed they are interacting with an AI, not a human being. This applies at the moment of first interaction and must be clear, accessible and not buried in terms and conditions. The exemption applies only where it is obvious from context that a system is automated.

For any company deploying customer-facing AI tools across EU markets, this has immediate practical implications. A chatbot must identify itself as AI in the language in which it is operating — whether that is Italian, German, French, Spanish or any other EU language — and the disclosure must be phrased in terms that are genuinely comprehensible to the user in that language. A technically correct but unnatural disclosure — the kind that results from machine translation without human review — risks failing the "clear and accessible" standard the European Commission's draft Guidelines of 8 May 2026 require. This applies equally to every target language: the obligation is language-neutral, but its execution is not.

2. Machine-readable marking of synthetic content

Providers of generative AI systems — systems that generate audio, images, video or text — must ensure that outputs carry machine-readable markers that identify the content as artificially generated. This is a technical obligation aimed at content provenance: enabling detection tools and downstream platforms to verify the origin of content.

The Commission interprets "synthetic content" broadly. AI-generated or AI-manipulated translations, summaries, image alterations and other modifications that materially affect the meaning or appearance of content fall within scope. This is not limited to sophisticated deepfakes. A translated product description generated by a machine translation engine and published without human review is, under this framework, synthetic content subject to marking requirements.

3. Deepfake labelling

Deployers of AI systems that generate or manipulate images, audio or video resembling real people, places, objects or events must clearly label the content as artificially generated or manipulated. Exemptions exist for artistic, creative, satirical and fictional works, provided a disclosure of the AI-generated nature is made in an appropriate manner.

4. Disclosure of AI-generated text for public information

Deployers using AI to generate or manipulate text published with the purpose of informing the public on matters of public interest must disclose that the text has been artificially generated or manipulated. The exemption applies where the text has undergone substantial human review by a natural person who takes editorial responsibility for the content.

This fourth obligation has particular relevance for institutional communications, press releases, regulatory submissions and any form of public-facing content that a company distributes across EU markets — regardless of the target language.

What this means for businesses using AI translation

The intersection of Article 50 with the use of machine translation is one of the most practically significant aspects of the new framework — and one of the most misunderstood. The core question every company must now answer is: does your AI-assisted translation workflow constitute "synthetic content" under Article 50, and if so, what does compliance require?

Pure machine translation: disclosure required

If a company uses a machine translation engine — whether a standalone tool like DeepL or a generative AI assistant like ChatGPT — to produce a translation that is then published without structured human review, that output is synthetic content under the AI Act framework. The company, as deployer, must either:

• label the content as AI-generated in a machine-readable format, or
• ensure it undergoes documented human review before publication.

In regulated contexts — pharmaceutical labelling, medical device Instructions for Use, legal documents, CE declarations of conformity — this creates an immediate compliance problem regardless of the target language. Whether the translation is into Italian, German, Polish, Romanian or any other EU language, the content categories that carry the highest regulatory risk are precisely those for which machine translation without expert human review was already inadvisable. The AI Act now makes this a legal matter, not merely a quality matter.

MTPE as the compliance solution

Machine Translation Post-Editing (MTPE) — the workflow in which machine-translated output is reviewed, corrected and approved by a qualified human linguist — addresses both the compliance question and the quality question simultaneously.

According to the European AI Office's interpretation of Article 50, the disclosure requirement does not apply where the AI-generated content has undergone comprehensive human review under the editorial responsibility of a qualified person. ISO 18587, the international standard for MTPE, defines what "comprehensive" means in practice: a documented review process, qualified post-editors with relevant subject-matter expertise, measurable quality outcomes and an auditable workflow.

An MTPE process certified to ISO 18587 thus achieves three things simultaneously: it produces higher-quality output than pure machine translation, in any target language; it complies with Article 50 without requiring content labelling; and it creates the audit trail that regulators and clients can request to verify compliance. To understand how EURECO structures its translation and technology workflows to meet these requirements, see our services and technologies page.

The liability question for companies using translation vendors

The AI Act places obligations not only on the companies that build AI systems but on the companies that deploy them — deployers. If a company commissions a translation from an agency that uses machine translation without documented post-editing, and publishes the result on its Italian website or in its Italian regulatory submission, the company is the deployer. The risk does not sit with the vendor. It sits with the company that made the publication decision.

This makes vendor selection a compliance matter. Before commissioning translations into any EU language for content that will be published to EU audiences, companies should verify that their translation partner discloses AI use transparently, operates a documented MTPE process, holds relevant ISO certifications, and can provide an audit trail of human review for each deliverable.

Sector-specific implications across EU markets

Pharmaceutical and life sciences

Pharmaceutical and medical device communications were already among the most strictly regulated in the EU before the AI Act. EMA guidelines for multilingual regulatory submissions require translated PIL, SmPC and IFU content in each member state's official language — meaning a single product launch may require compliant translations into ten or more EU languages. National agencies such as AIFA (Italy), BfArM (Germany), ANSM (France) and MHRA (UK) each impose their own submission standards on top of the EMA framework. MDR obligations for IFU content add a further layer. Article 50 now places an AI transparency requirement on top of all of these existing frameworks, across every language pair involved.

Practically, this means that any company submitting translated pharmaceutical documentation to any EU national authority must now be able to demonstrate not only the accuracy of the translation but also the nature of the process used to produce it — in every target language. Machine translation used without post-editing in a regulatory submission creates dual exposure: rejection by the national authority and potential sanctions under the AI Act. Our work in the chemical and pharmaceutical sector is structured precisely to provide the documentation chain that both frameworks require, across all required languages.

Legal and contractual documentation

Contracts, articles of association, employment agreements and court filings are, by definition, published with the purpose of informing parties about matters with legal effect — in whatever language they are drafted. An AI-generated translation of a contract submitted to a counterpart in any EU country without documented human review falls within the scope of Article 50's disclosure obligations, regardless of whether the target language is Italian, French, Dutch, Polish or any other.

The complexity compounds across legal traditions. Civil law concepts do not map directly onto common law equivalents, and the specific terminology of each national legal system — the Italian Codice Civile, the German BGB, the French Code civil — requires translators with documented legal expertise in that jurisdiction's language, not just general linguistic competence. A machine translation that produces a linguistically plausible rendering of a legal concept that does not exist in the target jurisdiction's law creates contractual risk that no disclosure label can mitigate. For our approach to legal document translation and the competencies required across jurisdictions, see our sector page.

Website and digital content

For companies entering any EU market through digital channels — e-commerce platforms, SaaS products, localised websites — Article 50 intersects with both the transparency obligations and each member state's consumer protection framework. AI-generated product descriptions, terms and conditions, and privacy notices that have not undergone documented human review must be labelled as AI-generated, in every language in which they are published.

In practice, consumers in any EU market interacting with a product page labelled as AI-generated will draw their own conclusions about the brand's investment in their language and market. The disclosure requirement is legally necessary; the reputational consequence of relying on it as a primary compliance strategy is a separate problem that multiplies with every additional language a company operates in. Our website and e-commerce localisation work operates entirely within human-reviewed workflows across all language pairs, ensuring compliance without disclosure labels.

Technical manuals and industrial documentation

CE-compliant technical documentation — user manuals, spare parts catalogues, installation guides — falls within a particularly complex intersection of the AI Act and existing product directives. The Machinery Directive (2006/42/CE) already requires documentation in the official language of each country where a product is placed on the market: a single product sold across France, Germany, Spain, Poland and Italy requires five separate compliant translations. When the Annex I obligations take effect in August 2027, AI systems embedded in regulated machinery will face the full framework. But the translated documentation accompanying that machinery is already subject to Article 50 from August 2026 — across every language version.

A machinery manufacturer that uses AI translation to produce manuals in any EU language and publishes them without documented post-editing is, from 2 August, a deployer of synthetic content without disclosure. Combined with the Machinery Directive's own language requirements, this creates a compliance exposure that affects the product's ability to reach each target market. For the full picture on technical translation for industrial manufacturers entering EU markets, read our article on automotive translation, EU regulations, technical manuals and homologation.

The human oversight principle: why it matters beyond compliance

The AI Act's exemption from disclosure obligations — where content has undergone documented human review by a responsible party — is not a loophole. It reflects a deliberate policy choice: the regulation is not trying to ban AI from translation workflows. It is trying to ensure that a human being with relevant expertise takes responsibility for the output before it reaches another person.

This principle aligns precisely with what professional translation has always required in high-stakes contexts. Regulated EU markets are, by definition, high-stakes contexts across every language pair: regulatory submissions to national medicines agencies, contracts governed by national civil codes, technical documentation accompanying products sold under EU product directives, financial disclosures subject to local securities law. In all of these areas, the human oversight that the AI Act now mandates was already the only acceptable practice. The regulation has not changed the professional standard — it has given it legal force, and extended that force across every EU language.

For companies that were already using qualified translation partners with documented MTPE workflows and ISO certifications, 2 August 2026 changes very little in practice. For companies that were using raw machine translation or low-cost platforms without structured review, the same date marks a significant compliance exposure. The distinction is not about the technology used — it is about whether a qualified human was responsible for what was published.

To understand how translation quality is measured and documented in a compliant workflow, read our article on mastering compliant translation, which covers the quality frameworks and certification standards that underpin AI Act-compliant translation in any language.

A practical checklist for companies operating across EU languages

If your company produces multilingual content for EU markets — whether through translation, localisation or AI-generated copy, in any target language — the following questions determine your Article 50 exposure:

• Do you use machine translation or generative AI in any multilingual workflow? If yes, identify each tool, each language pair and each content type it produces.
• Is any AI-produced translated content published without documented human review? If yes, that content — in every language — requires either disclosure labelling or a structured post-editing process before 2 August 2026.
• Can your translation vendor demonstrate an ISO 18587-aligned MTPE process? If not, ask for written disclosure of which tools are used and what human review takes place before delivery, for each language pair.
• Do your customer-facing AI systems (chatbots, virtual assistants) identify themselves as AI in the local language at the start of each interaction? If not, this is a direct Article 50(1) violation from 2 August — across every EU language in which those systems operate.
• Are your regulatory submissions, contracts or compliance documents in any EU language produced with documented human expert review? If not, they are potential synthetic content under Article 50, with compounded risk from sector-specific frameworks in each jurisdiction.

EURECO's translation and localisation workflows cover all major EU language pairs and are designed to ensure full compliance with Article 50 requirements across all content types and sectors — without requiring disclosure labels, because human expertise is integral to every deliverable. If you would like to assess your current multilingual workflow against the AI Act framework, contact us for a free consultation.