The EU has long treated multilingualism as a trade-off between democratic access and administrative speed. AI is weakening that trade-off, but it raises a harder question: if machines increasingly mediate what citizens read, who is responsible for what EU institutions say in each language? Brussels should answer with clear rules and its own public infrastructure.
On 11 September, the European Commission sent the proposed EU–India trade agreement to the Council only in English. Legally reviewed translations take time, so the Commission, hoping to sign before the end of the year, promised them only in time for the European Parliament’s vote. France has opposed the move, arguing that technical provisions can only be properly scrutinised once translated.
Because trade falls within the EU’s exclusive competence, national parliaments will not ratify the agreement. Their leverage is indirect: holding their own ministers to account before the Council acts. If the text exists only in English at that point, scrutiny is unequal before it has even begun.
This is less a new round in Europe’s long quarrel over languages than a question of democratic access. Brussels can go on working in English; the issue is whether Europeans must also scrutinise its decisions in English.
The EU has 24 official languages: citizens may write to its institutions in any of them, and its legislation must be published in all of them. In practice, English dominates, and with good reason. According to the latest Eurobarometer, 47 per cent of Europeans can hold a conversation in it. Brexit has, paradoxically, strengthened its claim to be Europe’s lingua franca, since English is now the mother tongue of only a small minority of EU residents and the national language of no large member state.
The Belgian philosopher Philippe Van Parijs argues that multilingual groups gravitate towards the “maximin” language, the one best known by whoever knows it least. In Europe, that is English. On this view, a supranational democracy needs a common forum, and a common forum needs a shared language.
The argument is strong, and restoring French would only swap one linguistic privilege for another. But Van Parijs describes a destination, not the present. Most Europeans still cannot hold a conversation in English, and conversation is the easy part. Above all, a lingua franca is how Europeans talk to one another. It need not be how their institutions talk to them.
Following a meeting in a second language is one thing; reading a trade agreement, understanding a legal right or acting on a public-health warning is another. In February 2020, as Covid-19 spread, the Commission initially published information about its response only in English. The Italian version arrived days later, after Maltese and Spanish, even though Italy was then Europe’s hardest-hit country. Citizens should not need fluency in a second language to understand a decision that binds them or information meant to protect them.
For decades, every additional language meant more translators, more coordination, and more time, so publishing first in English saved time. That constraint is now much weaker: the Commission’s eTranslation system already covers all 24 official languages and produces draft translations almost instantly.
But technology does not close the gap between a translation that is available and one an institution will stand behind. A fluent machine translation can drop a legal qualification, smooth over an ambiguity, or subtly drift between language versions. The Commission itself acknowledges as much: machine translations on its websites carry disclaimers, and legal revision still requires human expertise.
Yet the alternative to an official machine translation is not the absence of translation. Parliamentary staff, journalists, and citizens can paste an English-only document into a commercial chatbot, trading an authoritative text they may not fully understand for a readable one no public institution is responsible for. Nor do these systems behave the same in every language: a 2026 audit of leading AI models found that some gave materially different answers to politically sensitive questions depending on the language of the question. Translations will circulate anyway. The question is who takes responsibility.
That responsibility should lie with the EU. The French philosopher Barbara Cassin has urged us to think “not against but with machine translation”. For an institution, that means using automation where it helps, flagging uncertainty where it remains and returning to translators and lawyer-linguists when the text demands it. Two things are still missing: a legal framework for institutional machine translation, and a public infrastructure to deliver it.
The framework would set the rules. Documents intended for public scrutiny should appear in all 24 languages on the day of publication. If a human translation isn’t ready, the institution would issue its own machine translation, clearly labelled as provisional, linked to the original, and open to correction. Texts submitted for democratic approval would still need legally reviewed versions before any formal decision. The status of each version, and who is responsible for correcting it, would always be clear.
Under such rules, national parliaments could have started reading the EU–India agreement in their own languages on 11 September, while lawyer-linguists finished the authoritative versions in time for the Council’s decision.
The infrastructure would make the rules workable, and here the EU is unusually well placed. For almost seventy years, its institutions have produced parallel, legally scrutinised versions of their texts. What looked like the cost of multilingualism is also one of the best multilingual data sets in the world. Alongside eTranslation, the Commission’s Directorate-General for Translation released the EU Institutional LLM in July 2026, enhanced with that corpus to perform across all 24 official languages. Organisations based in the EU can already download it. What is missing is a public service built on these tools: open to national parliaments, journalists, and citizens, and backed by an institution that answers for its output.
India, the EU’s partner in this very agreement, already treats language technology as public infrastructure. Its publicly backed Bhashini platform provides AI-based translation across Indian languages, and its parliament is developing Sansad Bhashini to translate debates and documents.
The same principle could extend beyond written documents. During Poland’s Council presidency in 2025, AI dubbing made ministerial press conferences available in Polish, English, and French, with professional checks and clear disclosure of the synthetic voices. A commissioner could one day speak in one language while Europeans hear the same address in theirs, provided the synthetic version is identified and the original remains available.
None of this changes how Brussels works internally: negotiators can keep negotiating in English, officials drafting in it. Multilingual democracy does not require translating every working document by hand into 24 languages at once. It does require institutions to make their decisions intelligible to the people expected to live under them.
AI will not settle Europe’s language politics, but it removes much of the practical excuse for treating English as the default public language. The EU already has the technology, the translators and the lawyer-linguists. What it still needs is the decision to put them at the public’s service, under rules that make it answerable for what Europeans read in all 24 of its official languages. A Union that can answer to its citizens in every language has a stronger claim to unity than one that answers in only one.
Federico Rampinini
Federico Rampinini is an Adjunct Lecturer at the University of Parma. A historian of modern European philosophy, he has held research positions in Austria, Germany and Canada, and has published in the British Journal of Aesthetics and the Journal of the History of Philosophy.
