We came up with some simple coding exercises (about 20 minutes total to implement, max) and asked candidates to submit their responses when applying. Turns out one of the questions regularly causes hallucinated APIs in LLM responses, so we've been able to weed out a large percentage of cheaters who didn't even bother to test the code before submitting.

The other part is that you can absolutely tell during a live interview when someone is using an LLM to answer.