What does it mean to say that an AI understands a legal concept? Suppose a model can define good faith, identify relevant legislation and cases, translate the term into several languages and produce a convincing explanation of how lawyers use it. This would certainly demonstrate a considerable degree of legal knowledge.

A more difficult test arises when the concept moves between jurisdictions. Can the system explain how the role of good faith changes between legal systems? Can it distinguish a jurisdiction where the concept operates as a broad organising principle from one where its role is narrower? Can it identify which legal mechanisms perform comparable functions where there is no direct equivalent?

Legal concepts require broad contextual definitions because they exist within networks of rules, institutions, procedures and consequences. Understanding a concept therefore also involves knowing something about what it does and under what circumstances.

This becomes particularly important when AI crosses borders. A system may know two legal terms and correctly recognise that they are commonly translated as equivalents. If it does not also recognise that one carries rights or consequences absent from the other, however, its understanding remains incomplete for many practical purposes.

It can therefore be useful to think about different layers of legal knowledge. There is lexical knowledge concerning what a concept is called, definitional knowledge about what it means within a jurisdiction, contextual knowledge about how it operates and comparative knowledge about how it relates to concepts elsewhere.

Large language models already demonstrate impressive capabilities across all of these areas. The practical question is how reliably they can do so in specialist cross-border contexts where relevant distinctions may be subtle and consequential.

One approach is to make more of the underlying structure explicit. Rather than requiring a model to infer the complete legal relationship each time, it can be fed jurisdiction-specific definitions, contextual information, comparative mappings and identified limitations. Once created, this structured information can be reused across different queries and applications, providing a consistent legal foundation rather than requiring the same comparative analysis to be reconstructed each time.

This also has implications for consistency. Individual answers will naturally vary depending on the question and context, but they can draw on the same underlying analysis of how concepts relate across jurisdictions. Whether this amounts to genuine machine "understanding" is a much larger philosophical question. For practical legal AI, the more immediate issue is whether a system has access to enough structured legal context to distinguish situations that lawyers themselves regard as materially different.

This is likely to become increasingly important as legal AI moves beyond generating convincing legal language and is asked to perform substantive work across different legal systems. In that context, the ability to reuse structured comparative knowledge and apply it consistently may be just as important as the model's ability to reason about an individual question.