Test them on the actual work, whatever it is. Make multiple example questions or one big take home project that's similar to the real work they'd be doing, and allow them to solve it organically, as they would on the job. If they pass it with AI, cool, they managed to use it effectively. If they fail it with AI, that's on them. You don't need to go out of your way to purposely try to confuse AI OR the human. Working together is going to be the new norm and the AI is just another assistant or teammate.
IMO you're not testing "is this person super smart and able to recite leet code from memory without any internet access or reference materials". You're testing "is this person going to work well in our team and contribute meaningfully". You wouldn't disrespect your current coworkers by asking "how much AI did you use in this PR" if it otherwise works and passes tests and seems readable and maintainable, etc. (and if it doesn't, that's a code quality concern, not a question of human vs AI provenance).
The juniors should have the same ability to use the tooling available to the rest of us. If they choose to never understand AI output more deeply, well, the same could be said of any third party lib or Stack post. It's usually not a deal breaker in run of the mill business apps anyway. In fact there's an argument to be made for adopting good enough popular solutions vs reinventing the wheel and adding more tech debt for minimal gains.