I remember having to test exactly how a bunch of different predicates actually worked, when the documentation thereof was a complete lie and the predicates just evaluated to false instead of creating an error when fed data they supposedly accepted (but actually couldn't).
Note: i did work on a real product which had parser inside that had such style of repair implemented, and such approach is also quickly mentioned in the dragon book as well
A better point is that the error messages from many compilers can be very obscure. This was particularly true 10-15 years ago.
Clang does something like this. If it notices a simple enough syntax error that it can correct, it still generates the error message but it makes the correction internally and continues the compilation far enough to generate useful error messages about the rest of the code. But it doesn't output anything or change the code.
Of course it's malice.