Ma and colleagues’ Adaptive Workflow changes the order of human and AI decisions using estimated correctness likelihoods. When the estimated human likelihood is higher, the user gives an initial judgment before viewing AI advice. When the estimated AI likelihood is higher, advice appears immediately. The user always makes the final decision. In an income-classification experiment with 289 valid responses, this workflow improved final accuracy compared with an AI-confidence-only interface. That result is specific to the tested task. Estimated likelihoods can be wrong, and the authors identify possible confirmation bias after an initial judgment. Ma and colleagues (2023)