On Friday, 8 May 2026, I was part of an online panel together with Marija Todorova (Assistant Professor, Education University of Hong Kong) and Joss Moorkens (Associate Professor, Dublin City University) and chaired by Alison Rodriguez (Past President of FIT) as part of the Online Symposium of the Study Group on Language and the United Nations. We chose to discuss our respective angles on “Is AI Use Changing the Role of Translators? Implications for Education, Culture and Ethics”.
Both Marija and Joss addressed some of the impacts from teaching in higher education – with the latest cohort of students now not just digitally native, but seemingly AI native. In addition, they also mentioned aspects of their research which focuses on understanding some of the needs for translation in the marginalised society, in disaster alleviation in Malawi (e.g. in the aftermath of Cyclone Freddy in 2023) as well as in the case of Burmese migrants in Thailand (for whom machine translation plays an essential role in their daily life). In these settings, the role of translation is naturally far removed from my comfortable in-house setting in public administration in Vienna.
The responses below are based very closely on the responses I gave as part of the panel to the questions Alison asked. Most of my recent posts, both on this blog as well as on LinkedIn focus on on the importance of the human component in human-machine translation being that of an Expert in the Lead (the topic I spoke about at FIT XXIII in Geneva) as well as some of the cognitive issues that face translators in the age of (N)MT/GenAIAn umbrella term to cover machine translation using either (Neuro) machine translation or GenAI-based translation (i.e. using LLMs). More technologies. In addition, I very briefly touched on shadow translation as a governance issue, an area I am starting to put particular thought into. One particular issue is about the continuing accountability that professional translators have.
AR: Michael, You have also spoken about ‘unanticipated’ cognitive effects of AI; can you tell us about how, as a specialist translator, you use AI and its effects or changes you see in your work?
- As the only in-house translator at my employer, I have been able to use AI to help with the terminology extraction process – although this has been very reliant on using authoritative texts (e.g. EU Directives and Regulations) as a starting point, rather than giving the LLMA large language model (LLM) is a type of artificial intelligence (AI) algorithm that uses deep learning techniques and massively large data sets to understand, summarize, generate and predict new content. More free rein. In this regard it can prove useful, as I have been able to find ways to do proactive terminology work, rather than only reactive work.
- Continuing LLMA large language model (LLM) is a type of artificial intelligence (AI) algorithm that uses deep learning techniques and massively large data sets to understand, summarize, generate and predict new content. More hallucination, and a tendency to back translate rather than directly cite authoritative texts has however limited its usefulness. Even if I use AI, I still remain accountable, so I use it only where I know that I can rely on its output .
- Regarding how it has changed my work – previously I got a lot more “ad hoc” gist translation requests from colleagues, whereas for the nature of such requests, they have opted to “ask the LLM” rather than “trouble the linguist”. While this means less time spent handling very short but urgent pieces, and more of a focus on more substantial texts, those short texts were also very useful in terms of knowledge building. Additionally, by taking an AI-led approach there is the potential governance issue from an increase in shadow translation activity.
- Regarding the cognitive impacts, post-editing is not uniformly easier than translating from scratch — it depends heavily on text type, language pair, and AI quality. Usually it proves quite demanding to establish exactly what tools colleagues have used, and to establish whether they have used a detailed (persona) prompt to try to steer the LLMA large language model (LLM) is a type of artificial intelligence (AI) algorithm that uses deep learning techniques and massively large data sets to understand, summarize, generate and predict new content. More. Not knowing how a particular translation has been arrived at, and the original source of terminology used can really impact the cognitive load when post-editing.
- AI use redistributes cognitive effort but does not reduce it. There is a trend towards translations receiving texts that colleagues have “translated” using AI. The deceptive fluency of the output leads them to think that a text is close to final. Colleagues see it as requiring a minor edit, but I often have to check the fine nuances with them to identify the shortcomings of the LLMA large language model (LLM) is a type of artificial intelligence (AI) algorithm that uses deep learning techniques and massively large data sets to understand, summarize, generate and predict new content. More translation.
- Decision fatigue exists for the human translator: correcting mistakes they would not have made wears down translators. This is particularly the case for LLMsA large language model (LLM) is a type of artificial intelligence (AI) algorithm that uses deep learning techniques and massively large data sets to understand, summarize, generate and predict new content. More making terminology decisions based purely on statistical probability. In larger texts, I have noticed a deterioration of terminology consistency (likely due to limited context windows).
- Individual differences matter greatly, experienced translators manage AI-assisted workflows differently than novices. High-quality AI output can genuinely reduce load, but poor-quality output can increase it significantly compared to translating unaided – this boils down also to the deceptive fluency of AI output – and the fact that with many prompts by colleagues using a chatbot to serve up a translation, even if they try to use RAGRetrieval-augmented generation (RAG) is a technique that LLMs use to retrieve and incorporate new information from external data sources. The LLMs firstly refer to a specified set of documents and then respond to queries by the user, with these documents acting as supplementary information to the training data that has been used to train the LLM. In the early days of More (retrieval augmented generation), the LLMA large language model (LLM) is a type of artificial intelligence (AI) algorithm that uses deep learning techniques and massively large data sets to understand, summarize, generate and predict new content. More may choose to back translate rather than cite directly from the information it has been fed.
AR: AI was supposed to give us more time and mental space for creativity, but it has largely given us more to manage. The gains are real but narrowly distributed — often flowing to employers, platforms, and clients in the form of lower costs and faster delivery, while the human doing the work absorbs the new complexity, the new anxieties, and the new forms of invisible labour that AI oversight requires. Michael, can you share your professional experience and view on the possible overreliance on technology and its effects?
- All the tools involved lead to the underlying skill of translators being forgotten. The impact of tools trivialises these skills. Effective tool use is only part of the skillset of a professional translator. In AI literacy and competency terms, translators are “thrown in the deep end”. They receive only very general training, rather than dedicated training about how they really can benefit from AI.
- In terms of what clients expect/see/need, the old price/quality/speed triple constraint where “quality was king” no longer exists. Customers now expect near instantaneous delivery that does not cost a commensurate amount. What they need from the human involved is accountability, credibility and expertise – which “good enough” quality may not deliver.
- In terms of what responsible professional practice looks like, translators have to work towards ensuring an outcome (so in my case, ensuring the timely successful conclusion of a supervisory procedure and to ensure that there is no reputational loss).
- Institutional settings frequently treat translation as an afterthought (i.e. only sending finalised documentsfor translation), rather than central to achieving the outcome (i.e. translation work starting before finalisation of documents).
- Regarding practices that still build the right meta-cognitive skills needed vs metacognitive demands (i.e. is this text really correct/source trustworthy etc), in the case of junior translators, one way is a stepped approach to including tools in their workflows, so they really consider the text and learn to understand the subject matter, rather than just receiving MT/LLM output to correct, until they have built up subject matter expertise.
- Translators have an ethical responsibility: translators still maintain a responsibility to question the source text, check facts, figures in a critical manner that AI does not, and to check with the author. AI also leads to disrupted flow states, divided attention, decision fatigue, that make work feel more effort without the creative reward.
- Many translators are very familiar with their CAT environment, while AI tools draw them out of their state of flow, although the latest iterations of some tools are starting to draw AI tools into CAT flows to address this. Translators ultimately remain more committed to a translation where they actively tackle it, rather than just being a downstream verification instance.