When complex systems produce stable outputs, it is easy to assume that uncertainty has been resolved. In reality, though, it has often just been reassigned. Systems need stability to produce consistent output at scale, so categories have to be defined and boundaries drawn. The uncertainty surrounding those decisions does not disappear as a result, but moves upstream into label design, data structure and classification.

Consider a multilingual legal AI tool operating with a category such as termination. That category might encompass concepts relating to dismissal without notice, rescission following breach and statutory invalidation. At output level, the system may appear perfectly certain because the issue has been classified as termination. However, part of that certainty results from distinctions in the underlying material already having been absorbed into a broader category.

This is closely related to the problem of equivalence. If a system needs to treat different things as sufficiently similar in order to operate, an important part of its reliability lies in deciding when similarity is sufficient. From the user's perspective, these decisions are largely invisible because the user encounters the answer rather than the classification architecture behind it.

The issue becomes particularly important in cross-border legal AI because differences that appear minor linguistically can be significant legally. Scope, remedies, procedure and legal consequences may vary even where concepts look similar. A model can only work with distinctions available to it, so if uncertainty has been dealt with upstream by collapsing two concepts into a single category, the model may have no reliable way of reconstructing the distinction later.

Uncertainty cannot be entirely eliminated from law, and any system operating across numerous legal systems will require some degree of abstraction. The relevant questions are therefore where uncertainty is located, how it is represented and whether the people relying on the resulting output understand that it remains. As AI moves further into legal workflows, distinguishing stable output from genuine certainty is likely to become increasingly important.