Well, if by AI you mean large language models, they tend to do better at language tasks than math tasks. So a better example might be that it’s easier to get an LLM to write a statement for you and checking if it’s correct than writing the statement from the bottom.

The square root was just a clearer example. In the case of OP, it might very well be easier to have an LLM propose relevant case law and then check if that case law exists and is relevant, rather than having to find it yourself from square one.

It’s actually often easier to check an answer than coming up with an answer. Finding the square root of 66564 by hand isn’t easy, but checking if the answer is 257 is simple enough.

So, in principle, if the AI is better at guessing an answer than we are, it might still be useful. But it depends on the cost of guessing and the cost of checking.