Multilingual Legal AI Requires Data, Not Just Better Models

By Michael Krallmann, CEO, TransLegal.

For the legal tech community, cross-jurisdictional meaning raises questions of risk, liability and trust. Increasingly capable models, wrapped in sophisticated products, still encounter failures in the cross-border context. While the instinctive response is to look for a better model, the real constraint sits elsewhere.

Better models are not enough

Multilingual legal AI does not fail because models are insufficiently powerful. It fails because they lack access to data showing how legal concepts actually function across systems. Legal meaning is not universal, being the product of culture, religion, history, doctrine, procedure and institutional development. Without structured exposure to jurisdictional context and differences, even the strongest models default to the dominant framing in their training data.

The prevailing belief seems to be that scale and generalisation will resolve jurisdictional differences, that having enough text will enable a model to infer necessary distinctions. In practice, however, many of the most important differences are subtle, under-documented or expressed through practice rather than explicit definition. In other words, they require interpretation, not pattern matching.

Accurate data as legal infrastructure

A comparative law analysis, when done properly, is not merely a paraphrasing exercise. Rather, it identifies purpose, scope, legal effect and limitations, recognises partial equivalence and non-equivalence and, not least, states explicitly where an apparently correct translation breaks down. As such data do not emerge automatically from large corpora, they must be deliberately created.

This is why data architecture matters. A system designed to support cross-border legal work needs more than a large corpus of documents. It needs structured representations of legal concepts, mapped across jurisdictions, curated by experts. It also needs quality controls that reflect legal reasoning, not just linguistic correctness.

This kind of data-building is rarely done in a commercial context because it is slow, expensive and difficult to scale. Yet it is precisely such data that determine whether an AI system can be trusted in a multilingual legal environment.

There is growing recognition that wrappers alone cannot solve this problem. Prompting can improve presentation, and retrieval can improve relevance. However, neither solves the problem of conceptual (mis)alignment. If a system does not know that two terms only overlap partially, it cannot warn the user, and if it does not know that a familiar concept carries different implications and consequences in another jurisdiction, it cannot explain the risk this involves.

The alternative is to build datasets which define, contextualise and link legal concepts across jurisdictions. This entails embedding human judgement into the data itself, rather than relying on a model to infer it later on. It also means accepting that some answers need to specify uncertainty or limitations.

Reducing risk and extracting real value

For lawyers and legal tech builders, the problem of conceptual divergence reframes the role of AI. The most valuable systems will not be those that generate the longest or most fluent answers. They will be those that help users avoid mistakes they did not know even they were making. In cross-border work, that often means making differences visible rather than smoothing them away.

There are clear practical implications to this. Organisations that deploy legal AI across markets need to ask harder questions about the provenance and structure of the data their tools rely on. They need to consider whether the system offers transparency and accountability, and think about how legal risk travels when language crosses borders.

Legal AI will increasingly operate in environments where jurisdictional context matters. Systems that acknowledge this openly and are designed accordingly will age far better than those relying on linguistic gloss alone.

Towards a solution

For those interested in what a data-driven approach to cross-border legal accuracy looks like in practice, it is worth considering how structured legal terminology and comparative datasets can improve the output of AI systems. Engaging with the hard questions now is easier than doing so later once the costs of being almost right start to bite.

At TransLegal, we’re developing human-curated legal datasets aimed at improving AI performance in multilingual and multi-jurisdictional legal environments. Our work focuses on expert-led data creation and quality assurance to support more accurate and accountable legal AI systems. To date, we have completed over 40 jurisdiction-specific datasets, and have dozens more projects ongoing.

For teams building or integrating legal AI across markets, the question is no longer whether models perform well, but whether their base data support jurisdictional accuracy. To explore what this looks like in practice, visit TransLegal to try our data demo model or contact us directly.

About the author: Michael Krallmann is CEO of TransLegal, where he leads the development of structured cross-jurisdictional legal datasets designed to improve conceptual accuracy in multilingual legal AI systems. He holds a PhD in Law and Translation and works at the intersection of comparative law, language and legal technology.

–

[ This is a sponsored thought leadership article by TransLegal for Artificial Lawyer. ]