Crackpot Theory: Copilot (and by association many ML tools) is a form of probabilistic encryption. Once encoded, it's virtually impossible to pull the code (plaintext) directly out of the raw ML model (the cyphertext), yet when the proper key is input ('//sparse matrix transpose'), you get the relevant segment of the original function (the plaintext) back.

We've even seen this with stable diffusion image generation, where specific watermarks can be re-created (decrypted?) deterministically with the proper input.