Applying Bainbridge’s practice-loss mechanism to the two feedback levels yields a conditional hypothesis. If AI takes over routine search while scientists still judge the goal and evaluator, higher candidate scores do not settle those retained decisions. If the displaced work also supplied the practice and feedback needed to acquire that judgment, removing it could weaken capacity without removing responsibility. Investigate which judgments actually remain and how people can learn and exercise them; assigning authority alone is not a remedy. Bainbridge’s November 1983 industrial-control analysis supplies the historical mechanism, not a test of this transfer to AI-assisted science. The hypothesis implies neither universal human indispensability nor an inevitable increase in human work with more automation.