This still wouldn't be perfect of course - AIML101 tells me that if you get an ML model to perfectly respect a single signal you overfit and lose your generalisation. But it would still be a hell of a lot better than the current YOLO attitude the big labs have (where "you" is replaced with "your users")