Assuming they are trained on datasets that do not control for this classification, and are derived from the product of more than ten individuals I think it’s safe to expect, unironically, that yes the results would be more predictive. (ETA: but not grounded, and pointless as a survey.)