Arguably the tests should be easier to review than the implementation.

But if there non-trivial logic in the code of the tests, I agree this is probably a risky approach.