When Legal Terminology is Correct But the Answer is Still Wrong

By Michael Krallmann, CEO, TransLegal.

Legal AI is increasingly capable of producing outputs that look correct, not only linguistically, but also terminologically. Definitions align with established usage, translations follow accepted conventions and explanations reflect familiar legal structures. In cross-border work, this is often where the problem begins rather than ends.

In professional settings, correct terminology is usually taken as a reliable signal of accuracy. When a term is widely used and consistently translated, there is a natural tendency to assume that the underlying meaning travels with it. In law, that is a risky assumption. Legal concepts aren’t defined by their label alone, but also by their purpose, scope, conditions of application and, ultimately, their legal consequences.

Two concepts can align closely at the level of terminology and still diverge materially in practice, particularly across legal systems. One example is the relationship between liquidated damages in common law jurisdictions and contractual penalty clauses in many civil law systems. Translation-based systems will map one term to the other without hesitation, producing a result that reads cleanly and appears professionally sound. Yet in common law jurisdictions, liquidated damages are enforceable only if they represent a genuine pre-estimate of loss, whereas in many civil law systems, penalty clauses are generally enforceable even where they go beyond compensatory loss, subject in some cases to judicial adjustment. In other words, correct terminology is not the same as correct legal meaning, particularly where legal consequences diverge.

This would be manageable if legal terminology operated in isolation. In practice, however, it doesn’t. For a lawyer, a term carries a set of embedded assumptions about enforceability, remedies, procedural context and interpretation. When a familiar term appears, those assumptions are automatically activated. Readers fill in the gaps without necessarily realising that they are doing so.

This is where the real risk emerges. Wording may be correct but still invite the reader to apply an incorrect legal framework. The problem isn’t simply an imperfect translation. Rather, it is that the output triggers a line of reasoning that doesn’t hold in the target jurisdiction.

Errors of this kind are difficult to detect because texts look credible and precise, and follow patterns lawyers are trained to recognise. Under time pressure, or in routine workflows, it is easy for such output to pass review without deeper scrutiny.

The source of the problem sits in the data, not in how the output is reviewed. Foundation models are trained on large volumes of legal text, but not necessarily on structured representations of how legal concepts relate to one another across jurisdictions. They don’t carry explicit mappings of where concepts overlap, where they only partially align or where they lead to different legal outcomes. Faced with a gap, the model produces the most plausible approximation.

Prompting, interface design and retrieval can improve presentation and relevance, but don’t solve the problem. If a system doesn’t contain information about how two concepts differ in scope or effect, it has no basis on which to flag those differences. Output remains fluent, but the underlying legal position is incomplete.

This raises an uncomfortable point for lawyers and legaltech providers: legal AI can produce the right words and still point to the wrong answer. The closer the terminology appears to be, the easier it is to miss the gap.

Since the problem sits in the data, it has to be addressed there. Legal meaning needs to be represented explicitly, with attention to purpose, scope and legal effect, and with clear identification of where concepts don’t align cleanly. Such information doesn’t emerge automatically from large corpora. It has to be constructed, curated and maintained.

At TransLegal, we build structured, human-curated legal datasets that map concepts across jurisdictions and capture relevant distinctions directly in the data. We also use AI systems to generate additional layers of comparative legal data, with experts in the loop to review, validate and refine the output. This extensive data can then be used to help AI systems surface differences rather than smooth them away, and to give users the information they need to recognise where apparent equivalence breaks down.

As legal AI becomes more widely used in cross-border contexts, the quality of terminology will matter less than what sits behind it. Outputs can look entirely correct and still lead the user in the wrong direction. This problem sits directly in the path of cross-border legal work and is already shaping outcomes.

—

To explore how TransLegal can help you, see here, or you can also contact the team 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. ]