Much of the discussion around multilingual legal AI understandably begins with technology. Another useful starting point is multilingual law, where legal systems have been trying to communicate authoritative legal meaning across languages for a very long time.

EU law provides perhaps the clearest example. One of its more unusual features is that all official language versions are treated as equally authentic. This principle is important because citizens and businesses should be able to rely on EU law in their own language rather than having to refer to another language, which they may not speak, as the authoritative version.

In practice, however, equal authenticity does not mean that every language version always points towards exactly the same interpretation. Linguistic formulations can diverge, sometimes immaterially and sometimes in ways that matter materially to the legal analysis. The Court of Justice of the European Union has therefore long had to compare different language versions and interpret them by reference to the wider context and purpose of the legislation.

The broader point is that legal meaning does not always travel cleanly between languages. Words exist within linguistic and legal systems, and legal concepts carry assumptions about scope, purpose, application and legal consequences. Different languages and legal traditions may express those assumptions differently.

This problem becomes even more pronounced when the analysis moves beyond different language versions of the same legal instrument and across different jurisdictions. A term such as good faith may be familiar in several legal systems while its doctrinal significance varies considerably. Likewise, contractual penalty mechanisms can exist across jurisdictions while the rules governing their enforceability differ. Familiar terminology creates a useful bridge, but does not by itself establish legal equivalence.

Multilingual legal systems have developed ways of dealing with these tensions. They do not assume that every divergence can be eliminated. Instead, differences may need to be identified, compared and interpreted. This is an important lesson for legal AI.

Artificial intelligence changes the scale at which multilingual legal material can be processed. A system can now translate, compare and generate legal content across languages and jurisdictions at a speed that would previously have been impossible. This creates obvious opportunities, while also allowing assumptions about linguistic and conceptual equivalence to propagate just as quickly.

An AI system encountering apparently corresponding terms may reasonably infer that they express the same legal concept. In many cases that will be a good approximation. In others, however, the concepts may only partially overlap or lead to different legal consequences. Increasingly capable models will infer many of these distinctions from context, but cross-border reliability should not depend entirely on whether the relevant relationship is reconstructed correctly every time a question is asked.

Important relationships can also be made explicit. Concepts can be defined within their own jurisdictions, their scope and legal effects recorded and their relationship with concepts elsewhere analysed. This gives downstream AI systems more information on which to base their reasoning and creates the possibility of preserving meaningful divergence rather than smoothing it away.

The experience of multilingual law therefore offers a useful principle for legal AI. The objective should not be to make different languages and legal systems appear more uniform than they really are, but to enable technology to work across them while retaining enough structure to recognise when the differences matter.