A high-risk AI system may produce the same output in several countries. The people interpreting that output still need instructions they can understand and act on.
In brief
- For high-risk AI systems, Article 13 of Regulation (EU) 2024/1689 requires instructions for use that are concise, complete, correct, clear and comprehensible to deployers.
- Recital 72 explains that the instructions should be available in a language easily understood by target deployers, as determined by the Member State concerned.
- Translation must preserve limitations, uncertainty qualifiers and escalation steps, and stay aligned with the interface.
The operative rule and its audience
Regulation (EU) 2024/1689, the AI Act, has phased application dates. For high-risk systems, Article 13 requires instructions for use in an appropriate format that are concise, complete, correct, clear, relevant, accessible and comprehensible to deployers.
They must describe the intended purpose, performance and limitations, foreseeable risks, human oversight measures and other information needed for proper use. Recital 72 explains that the instructions should be available in a language easily understood by target deployers, as determined by the Member State concerned.
These provisions are about high-risk AI systems and their deployers. They do not create a blanket duty to translate every AI interface or marketing page into all EU languages. Teams should first classify the use case, identify the relevant duties and confirm the applicable start date; the regulatory owner makes that call.
What can go wrong in translation
An instruction may tell a human reviewer when to disregard a model score, which input data is unsuitable, or when to escalate a decision. If a limitation is softened or an uncertainty qualifier disappears, the person supervising the system may over-rely on the output. Translation must preserve the boundaries of intended use and the exact relationship between a metric and its testing conditions.
Language work also spans a release system: instructions, in-product help, training, oversight procedures and change notices must describe the same version. A screenshot or screen label that diverges from the instructions creates a practical failure even when the PDF itself is accurate.
A brief for product and compliance teams
- Define the deployer. Role, sector, training and likely language of day-to-day operation.
- Supply structured source content. Intended use, exclusions, performance limits, input requirements and escalation steps.
- Build a terminology set. Keep “output,” “confidence,” “review” and “override” consistent in context.
- Test the workflow. Review instructions alongside interface strings and realistic scenarios with the responsible subject experts.
- Track changes. Recheck translations when a model, threshold or oversight measure changes.
Good instructions help the person in control notice when the system is outside its limits. That is the language outcome worth designing for.
Related reading: how we support AI teams and compliance leaders. The European Commission’s AI Act page ↗ sets out the application timeline, and the AI Act Service Desk ↗ offers interactive tools to help organizations check whether they are subject to its obligations.
Primary legal text: Regulation (EU) 2024/1689 ↗ (Article 13 and recital 72). Further reading: European Commission: AI Act ↗, AI Act Service Desk ↗. This article explains selected language provisions and does not replace a product-specific regulatory assessment.