In 2024, the European Union adopted the Artificial Intelligence Regulation (AI Act), which concerns the use of artificial intelligence systems in various fields. The European Commission (EC) has now drafted guidelines to clarify how the Regulation’s individual provisions should be applied in practice. One of these guidelines addresses the so-called high-risk artificial intelligence systems, which are systems whose use may significantly impact individuals’ rights. Prior to the final adoption of the guidelines, the EC launched a public consultation, inviting experts, organisations and individuals to participate.

Why is this relevant to translators and interpreters?

At Association of Translators and Interpreters of Slovenia (DPTS), we are concerned that the AI Act does not categorise AI-generated translation and interpreting as high-risk activities. While we know that the quality of software products in this field is improving, we also recognise that their accuracy level is not on par with that of human translators and interpreters. We strongly believe that in sensitive areas, such as law, administrative proceedings, medicine, law enforcement and asylum proceedings, the focus should remain on the risks involved. Even the slightest inaccuracy in translation or interpreting can cause great harm, negatively impact fair trials, result in situations where the rule of law is not upheld and all parties do not have full access to justice.

When used for translation, AI systems struggle with complex contexts, lack consistency, and are not sufficiently accurate. Sometimes they change affirmative statements to negative ones, or vice versa. They may add some phrases (hallucinations or confabulations) or omit crucial parts of a sentence or an utterance, have problems with slang and dialects, and fail to understand the emotional context in which an utterance is made. In the field of interpreting, the risks associated with machine processing of human speech include unnatural prosody, poor sentence segmentation and unnatural pacing. There is also a risk of failure to grasp the contextual and emotional integrity of sensitive narratives. AI systems may also struggle to understand speakers with speech impairments or speech affected by strong emotions such as crying, fear, stress, shock or panic, which can significantly affect speech production and intelligibility.

As language professionals at DPTS, we therefore think that it is of utmost importance that the EC guidelines on AI Act risks clearly state in Annex III, points (6), (7) and (8) that AI-generated translation and interpreting are risk-involving activities.

What did DPTS do?

The Executive Committee of DPTS decided to play an active role in the EC consultation process. An external expert in artificial intelligence law (Gregor Strojin of Deloitte) was commissioned to carry out a legal analysis of the draft guidelines. The ExCom working group then drew up comments from the perspective of translators and interpreters. In addition to the association’s joint contribution, members of the association’s governing bodies were invited to participate as individuals and submit their own comments, if they wished.

What did we propose?

In our comments, we emphasised that the guidelines should pay closer attention to the use of AI in translation and interpreting. Specifically, we proposed the following:

1) that the concept of ‘autonomy’ as used in the Regulation be interpreted broadly, to include the use of AI tools for translation and interpretation in law enforcement, asylum and cross-border proceedings, administrative and judicial proceedings, and democratic processes within the scope of high-risk activities. Artificial intelligence can translate a statement, produce a transcript of a hearing or summarise an interview independently. If such a document becomes part of the official record, any errors may also affect subsequent stages of the proceedings, as later decisions are based on it. This technology therefore influences important decisions affecting people and raises serious questions regarding the protection of personal data and confidentiality. It is therefore a tool involving unacceptable risks, which must be strictly restricted by European legislation;

2) that the intended purpose of the AI use be taken into account. If a translation, transcription or summary becomes a part of the official record, it may affect the understanding of the facts of the case and, consequently, the decisions made during the proceedings, even if these tasks are considered ‘preparatory’. An incorrect translation may be carried over into subsequent stages of the proceedings, where key decisions are based on it.

3) We encourage the AI Office to draw on human rights risk assessment methodologies (e.g. HUDERIA), which assess risks to human rights, democracy and the rule of law in a structured way. Applying these methods to AI translation and interpretation tools shows a significant likelihood and severity of harm in Annex III, points (6), (7) and (8). This supports their classification as high risk due to a high likelihood of error and severe impact on fair-trial, non-discrimination and privacy rights. The same AI translation or interpretation tool can pose a low risk in everyday use and a high risk in contexts covered by Annex III. Similarly, risk levels may differ significantly based on the language pairs involved. The guidelines should therefore treat AI‑based translation and interpreting as a paradigmatic example of how context determines risk.

4) We therefore propose that the guidelines explicitly recognise AI‑based translation and interpretation when used in law enforcement as high risk under Annex III (6). When used in border screening, asylum interviews, detention, and return/removal procedures, they must be recognised as high‑risk under Annex III (7), and when used in court proceedings they must be recognised as high‑risk under Annex III (8)(a).

5) We further propose that the eventual use of AI-based translation and interpretation should be transparent and made available to procurers, deployers and end users of such technologies through appropriate technical documentation, including accuracy levels for specific language pairs and error rates.

At DPTS, we believe that the EC consultation on the draft guidelines has presented an opportunity to draw stakeholders’ attention to practical dilemmas and unclear cases of application that could affect how the rules will be applied in practice. At the same time, the fact remains that AI can only ever be an auxiliary tool. The real work must be left to experts who understand not only languages, but also people, cultures and the context in which we communicate.

DPTS will continue to monitor developments in European AI legislation and strive to ensure that the specific characteristics of the translation and interpreting professions be taken into account in its formulation. While the development of AI brings new opportunities, it also raises important professional and ethical questions, in which we, as a professional association, wish to play an active role.

Once the EC has published the responses to its consultation, we will post the relevant link on the DPTS website.