I agree -- visually judging outliers is a pretty poor test of statistical significance. The fact that they sat near each other does give them some weight though.

The big problem is that it only works for people who share a fairly unique answer set. In other words, they need to all have a decent number of wrong answers which correlate. What about people at the bottom-right of the graph who had a high correlation simply because they answered nearly all of the questions correctly? Certainly most of them are innocent, but how would we know? It seems this technique can only finger people who cheat and still end up with a mediocre grade, not those who cheat and end up acing the exam.