AI models have become remarkably good at producing fluent multilingual content. In legal work, however, fluency only addresses part of the problem because a translated legal term can be linguistically correct while still creating the wrong legal impression.

This is particularly difficult for users who do not know the target jurisdiction or language well enough to recognise the problem themselves. If an AI-generated translation looks professional and uses established terminology, there may be little reason for the user to question it. Human review offers one solution, and local lawyers and lawyer-linguists can identify hallucinations, misleading near-equivalents and inappropriate terminology. However, repeated expert review also consumes time and reduces some of the efficiency gains AI promises.

One way of addressing this is to move some of that expertise upstream. Cross-border legal work requires more than one-to-one terminology; it also requires information about how a concept functions in the relevant jurisdiction, including its scope, purpose, legal effects and limitations.

Consider liquidated damages. A superficially corresponding mechanism may exist in another legal system while being governed by different rules concerning enforceability or judicial intervention. Translating the label correctly does not communicate those differences. General-purpose language models can produce plausible language from patterns in their data, but comparative legal relationships are not necessarily available to them as explicit, structured information.

A model asked for the closest equivalent of a legal concept therefore needs more than another term. Ideally, it should have access to information explaining whether concepts are genuinely equivalent, substantially similar, partially overlapping or materially different. Structured multilingual legal data can provide this type of anchor by supplying jurisdiction-specific definitions, comparative context and information about degrees of equivalence when an application needs them.

At TransLegal, this is the main infrastructure problem we are working on: how to construct structured legal information that allows downstream AI applications to recognise jurisdictional differences rather than having to infer them from linguistic similarity alone. Model capability will clearly continue to improve, but for cross-border applications reliability will also depend on the quality and structure of the legal information those models can access.