[jabbany]

I wonder if OCR could be improved by adding a "language model" of sorts...

Like, sure, maybe it's hard to tell apart a "1", "i", or "l" purely visually, but if you knew it was supposed to be code, I'd suspect one could significantly improve the recognition accuracy if the system just worked in the probability of each confusable option given the preceding (and following) text.

[playingalong]

This would also have a higher risk of introducing some nasty, hard to spot errors.

It's actually better for the compilation to fail than for the Clippy to make up something syntactically and compilation correct, but wrong.

[eru]

You might be right in a practical sense, but for an art project like this, maybe not?