The solution for this is property based tests. Outputs almost never have maximum entropy. They have invariants, and they're related to the inputs in various ways.

Trivial example: authenticated encryption. The input is some random text, some random key, and some random nonce. The result is a message. One property to check is the fact that the message can successfully be decrypted with the same key, and the resulting plaintext will be the same as the random text from the beginning. Another property is that any message that is not the message you computed will fail to decrypt (assuming you pretend not to know the key when as you corrupt the message). Testing (crudely) that you can detect forgeries (or at least accidental corruptions).