That sounds like a fun application of a neural network. Finding large repositories of "real" papers to train it with should be fairly easy, too.

That said, I have zero experience working with neural networks, so it might not be very easy at all.

You'd get pretty far just by whitelisting arxiv, .edu, .ac.uk, etc.