I don't see why. The code could be very complicated to understand, or the grading mechanism could be hard or impossible to reverse. (Just like cryptographic hashes.)

Better would be access to the training corpus. Then use that to train, say, a Markov generator, and have it replace words from a thesaurus, in hopes of bypassing a similarity detector.