Using a tool like this to hide sections of code presented for review places a lot of trust in the automation. If Mallory were to discover a blind spot in the semantic diff logic, she could slip in a small change for eventual use in an exploit, and it would never be seen by another human.

For example, consider this part of the exploit used in the recent xz backdoor. In case you don’t see the problem, here’s the fix.

Rather than hiding code from review, if a tool figured out a way to use semantic understanding to highlight code that might be overlooked by a human (and should therefore be reviewed more carefully), it could conceivably help find such things.