Ask HN: How do you differentiate with AI coding interviews?

I haven’t interviewed for a coding job in a while, not since before AI coding was a thing.

I’m wondering: if your interview process allows the use of tools like Claude and Codex, how do you differentiate candidates?

When I was doing interviews, you would be given some sort of problem statement, and 20 to 50 minutes typically to solve JT. Sometimes you might get a take home problem, or multiple hours to do it. The basic idea was always the same: write some code, and the interviewers judge the quality of your code.

There was a lot of debate about the fairness of these processes, but you couldn’t deny that they differentiated candidates. On any problem, some candidates did well and some did bad. They all produced different results.

With AI tools, all candidates presumably put in the same prompt (the interview problem statement). If you ask them follow up questions, they all presumably can ask that question to the same LLM.

So, beyond some basic floor, do interviews still differentiate candidates? How?