Everything is flawed. But if you're testing the effect of "AI advice" you need to use an AI that has a normal failure rate without tailoring the questions to alter that rate, and/or compare AI versus other sources of information that have the same failure rate.