Translator at the International Monetary Fund: “AI Thinks in English – The World Doesn’t”

Written by: Natalie Foster

---

In 2025, the World Economic Forum highlighted a growing but often overlooked problem, generative AI systems remain heavily biased toward English. Trained predominantly on English-language data, large language models frequently struggle with nuance, register, and conceptual accuracy in other languages, a limitation that becomes particularly risky when AI is used for official, legal, and economic documents.

“AI has no idea what it is translating. Words mean nothing to it, it simply rearranges them based on patterns,” says Denis Pshenichnikov, a professional translator at the International Monetary Fund and winner of the UN St. Jerome International Translation Contest, one of the most prestigious international competitions in the field.

Pshenichnikov has spent years working with high-stakes multilingual texts at the IMF and the United Nations, coordinating large translation projects and contributing to the development of institutional terminology. His professional performance at the IMF has been highly rated, which implies little room for ambiguity. He is also the author of academic research on machine translation post-editing, where he examines how human expertise becomes more critical in an AI-assisted workflow.

In this interview with BrainZ, he reflects on his own experience working with AI-assisted translation inside global institutions, explaining why language imbalance in AI is not a technical glitch but a structural issue that he confronts daily in international communication and decision-making.

At the World Economic Forum in 2025, experts warned that generative AI is still heavily biased toward English. From your experience working with IMF and UN documents, how visible is this bias when AI systems are applied to multilingual, high-stakes texts?

Depends on the language. There are several machine translation tools and some work better with some languages and others with other languages. Also, for some languages like Spanish and Portuguese regional variations are significant and must be taken into account.

Another issue is register. AI frequently struggles to distinguish between formal institutional language and conversational tone. In official documents, especially in organizations like the IMF or the UN, wording must be precise and consistent. Many English terms have multiple possible equivalents in other languages, and AI does not always choose the correct one in context. For official documents issued by international organizations such as the IMF and UN human-supervised translation is essential.

You won first place at the UN St. Jerome International Translation Contest, where precision and nuance are decisive. What kinds of meaning or intent do AI systems most often lose when translating from or into non-English languages?

AI has no idea what it translates. The words mean nothing to it. It just tries to put the words in what it thinks is the right order based on the texts it was trained on and it does it extremely fast. An AI-generated translation can be a useful first draft which then needs to be edited by a professional translator. It’s especially important for more creative texts such as mass media publications or literary works where how you say it is often as important as what you say. AI does not understand different writing styles so official documents may sound too conversational and more informal texts, too formal. Another thing is grammar, especially for languages like Russian where grammar is very different from English. AI may use all the correct words, but it just doesn’t sound right. 

In global organizations, a single wording choice can influence how policies are interpreted across countries. Based on your work coordinating large translation projects at the International Monetary Fund, where does AI pose the greatest risk: vocabulary, tone, or conceptual framing?

All of those are important. Vocabulary, obviously, for texts where precision is needed such as financial and technical texts and as we know a lot of words can be translated in more than one way. The right tone is important for official statements and speeches where it’s necessary to make sure the translation conveys exactly the same idea the author had in mind. English is a compact language so sometimes you use five or six words to translate what English expresses in two words. 

In your academic research on machine translation post-editing, you argue that human expertise becomes more important with AI. How does language imbalance in AI models increase the cognitive and ethical responsibility of human editors?

Like I said, AI does not understand what it’s translating but people do. That fundamental difference becomes even more critical when the system is trained on uneven linguistic data.

One issue, which has also been noted in broader AI research, is that fluency can create an illusion of correctness. A translation may read smoothly and confidently but still contain conceptual errors or subtle distortions. Because generative models are optimized for plausibility rather than factual or contextual accuracy, the burden shifts to the human editor to detect what “sounds right” but is actually wrong.

From a cognitive perspective, this requires a different kind of attention. Translators must slow down, compare source and target texts carefully, and resist the tendency to trust fluent output. Post-editing AI-generated content is not the same as editing human translation. It requires awareness of typical AI patterns, like overgeneralization, stylistic inconsistency, or invented terminology.

You’ve worked on documents intended for governments with very different cultural and economic contexts. How does language bias in AI affect cross-cultural understanding, not just technical accuracy?

For some languages there are significant regional variations, for example Spanish in Spain and Spanish in Latin America or European and Brazilian Portuguese, and the same words have very different meanings. For other languages like Russian or Chinese the grammar differs significantly from English so often you must completely change the sentence structure.

When AI-generated or AI-assisted translations are used in official documents, who should ultimately be accountable for meaning the system, the institution, or the human expert? How is this responsibility handled in organizations like the UN and IMF?

The institution is responsible for all public documents it produces. The IMF has a rule that all official publications must be reviewed by a professional translator who is responsible for the final translation. There are various levels of MTPE for different kinds of documents, from light to full post-editing. The most thorough post-editing is given to official Fund publications either in print or on the website.

Your performance at the IMF has been highly rated which implies near-zero tolerance for ambiguity. Do you think current generative AI systems are compatible with such standards, or do they fundamentally operate by different logic?

Think of AI as a baby that is still learning and has yet a lot to learn but it’s learning fast. We can’t say for sure that it will ever be able to produce content equal to human output. At least for now it looks like AI has a long way to go before it can be used unsupervised, if ever. Also, at least presently, AI is learning based on what humans feed it and humans are not perfect and make mistakes so it’s inevitable that AI makes mistakes. Thus, a second pair of (human) eyes is essential. 

English dominates not only AI training data, but also global discourse itself. Do you see a risk that AI will unintentionally reinforce a kind of linguistic power imbalance in international decision-making?

It’s a natural process. English was not always the dominant language in international relations and will not always be. For now, this is what we are dealing with, and people have very little, if any, control over that process. As long as AI uses human knowledge it will reflect the changes in the human world.

The concern is that AI could amplify this imbalance. If most high-quality AI tools work best in English, and if decision-makers increasingly rely on AI-generated summaries, drafts, or analyses, then perspectives expressed in other languages may be filtered through an English-centric lens. This does not necessarily happen intentionally, but structurally.

Looking ahead, what would a responsible use of generative AI in multilingual global institutions actually look like and what role should human experts play in that system?

AI is and must remain a tool that helps perform a task but cannot act on its own and be held accountable for the final output. Responsible use requires human supervision. For professionals it means adapting by developing new skills. For example, we can argue that translators that use machine translation act more like editors because the first draft is already produced by the machine. Post-editing machine translation appears somewhat different than editing human translation and this phenomenon is currently extensively researched. Translators will still be needed but in a different capacity, more like editors and perhaps technical experts. 

Disclaimer: The views expressed herein are those of the author and should not be attributed to the IMF, its Executive Board, or its management.