If ai watermarking is undetectable to humans I wonder if sinister stuff in the context is also undetectable... Some thought or mood that you can't read but is still encoded in the tokens.