Is AI translation compliant with the MDR/IVDR for medical devices?

While some auditors may still take a different view, the fact is that AI can be used for translation within the regulatory framework established by the EU MDR/IVDR. In fact, this has been the case since 2017, when machine translation had already reached quality levels similar to those seen today with the large language models popularised by ChatGPT.

The use of artificial intelligence in translation is not only compliant within the context of Regulation (EU) 2017/745 (MDR) and Regulation (EU) 2017/746 (IVDR), but, when properly implemented, it may also be recommended for achieving more efficient, scalable and consistent processes.

Neither the MDR nor the IVDR prohibit the use of machine translation or AI technologies. On the contrary, the regulatory framework is outcome-focused, requiring that the information supplied with the product is safe, understandable and suitable for its intended use (Article 10 MDR/IVDR and Annex I, Section 23 MDR/Section 20 IVDR).

From this perspective, AI-based translation is compliant with the MDR/IVDR provided that it is integrated into a quality management system that ensures the conformity of the final result.

The actual requirements of the MDR/IVDR: clarity, official languages, accuracy and information safety

The MDR and the IVDR impose clear obligations regarding the information supplied with the product, including labelling, instructions for use and materials for the intended user. Most importantly, Member States may require specific languages, meaning that translation is not ‘optional’, but an inherent requirement for access to the European market.

In practice, this means two things. First, if you translate, you must translate well, as a defective translation can directly affect the safe use of the product. Second, it makes no difference whether AI, traditional machine translation or human translation from scratch is used; what counts is that the final result is correct, consistent and verifiable.

If you market your products in several European countries, it may be worth consulting the overview of language requirements published by the European Commission, which helps anticipate which documents need to be translated and into which languages. It is neither complete nor entirely up to date, but it is still a reasonably useful resource as a starting point. Although limited in number, there are some country-specific exceptions where translation is not required.

Finally, it is worth noting that the only specific reference in the regulations as to how translation should be performed can be found in Article 16 of the MDR/IVDR, paragraph 3: ‘…procedures which ensure that the translation of information is accurate and up-to-date…’, a requirement explicitly placed on importers and distributors. Given that manufacturers are ultimately legally responsible for translation, it can be understood that this requirement for translation accuracy also applies to them.

The regulatory key to translation: manufacturer responsibility for the final content

Article 10 of the MDR and the IVDR establishes that the manufacturer is responsible for the conformity of the product, including all accompanying information. This implies that conformity does not depend on whether AI has been used or not, but on whether the manufacturer can demonstrate that the final translation meets the requirements of clarity, comprehensibility and the applicable language requirements.

In other words, if in an audit you are asked whether you use AI for translation, the safest answer is not simply ‘yes’ or ‘no’, but rather something along the lines of ‘we have a controlled, validated and traceable process, with qualified human review and documented evidence of approval’.

Translation quality does not depend solely on the use of AI, but on how it is configured

AI-assisted translation does not constitute a single or standardised process. Its quality can vary very significantly depending on how it is configured, supervised and integrated into the manufacturer's linguistic and documentation workflow.

There are measures that steer the use of AI towards higher-quality and more regulatory-appropriate results, for example, through customised AI prompts according to the document type, the integration of glossaries and translation memories or the use of the most advanced models.

For this reason, the analysis should not focus solely on whether AI is used, but on whether it is used within a controlled, parameterised and verifiable environment.

Human review and control: an essential requirement for the use of AI with medical devices

For the use of AI in translation to be justifiable from a regulatory point of view, its results must be reviewed and terminologically verified against the reference documentation by qualified experts. Furthermore, the use of AI must be documented within the quality management system.

In this sense, AI does not remove controls; instead, it demands a more robust process design. The greater the level of customisation, terminology control and use of validated materials, the lower the risk associated with the use of AI in translation and the stronger its defensibility within the quality management system.

It is also worth considering a key point from the perspective of ISO 13485 for medical devices: a translation process based exclusively on AI would not be justifiable under clause 7.5.6, which stipulates that ‘validation must demonstrate the capability of these processes to consistently deliver planned results’.

Given the current state of the art, generative AI systems can produce non-deterministic results and show variability or errors that are not always detectable through indirect controls, which means they cannot, on their own, be considered compliant for the translation of labelling, IFUs, warnings, contraindications, etc.

The clear practical consequence is that validation cannot be based on AI as a translation tool, but rather on the control system surrounding it: qualified human review, terminological and documentary control and traceable approval records.

What evidence tends to works best in an audit?

In audits, what stands up best is not a ‘tool’, but a system that leaves a trail. It usually helps to have a documented procedure for the translation workflow with formal quality criteria, records of human review and approval and terminology control. In any case, remember that your QMS must include a procedure covering translation, as required by the MDR/IVDR.

A good test is to ask yourself: if an external party were to question the translation of a critical sentence tomorrow, could you identify who translated it, who reviewed it and what criteria were followed to approve that version?

In this regard, the issuance of a translation certificate is highly useful, as it can serve as a record in the quality management system. The certificate may include information such as translated documents, people involved in the process, date of issue of the certificate, languages translated and other information that may be relevant depending on the procedure defined in your own QMS.