The key bit is constructing a hard, novel proof. The fact that AI doesn't (yet) do this isn't evidence that it doesn't reason, but if it did so it would be strong evidence that it does reason.

(I also take the pessimistic point of view that most humans don't reason, so YMMV.)