More often than not the code in research needs to be flexible (from pbp to aggregate level). If AI can get you answers but a) uses functions no one has seen before, b) you can't explain what's happening, and c) adjusting the code is more than a few line changes, what's the point?
When we've giving coding tests in the past, pre-AI, we let the candidate do web searches all they wanted... because that's what we do in our regular job
Doing stuff by memory is silly in a brand-new environment
No idea about the AI-age