‘Discrimination’ can mean all sorts of things. One of the main ones, and what it will mean in this post, is differential treatment of people from different groups, due to real or imagined differences in average group features. Discrimination is a problem because the many people who don’t have the supposed average features of a group they are part of are misconstrued as having them, and offered inappropriate treatment and opportunities as a result. For instance a capable and trustworthy middle aged man may miss out on a babysitting job for which he is truly the best candidate because the parents take his demographic information as reason not to trust him with their children.
This means that ‘discrimination’ is really a misnomer; this problem is due to lack of discrimination. In particular lack of discrimination between members of the groups. For instance if everyone could instantly discriminate between women with different levels of engineering ability, generalizations would be useless, assuming engineering ability is really the issue of interest to the discriminators. Generalizations aren’t even offensive when enough discrimination is possible. Telling a 6’5” Asian man that he’s probably short since he’s Asian is an ineffective and confusing insult. Even if observers can’t discriminate perfectly, more ability to discriminate means less misrepresentation. For instance a test score doesn’t perfectly determine people’s abilities at engineering, but it is much more accurate than judging by their gender. This is assuming the generalizations have some degree of accuracy, if they are arbitrary it doesn’t make much difference whether you use false generalizations of larger groups or smaller ones.
The usual solution suggested for ‘discrimination’ is for everyone to forget about groups and act only on any specific evidence they have about individuals. Implicitly this advice is to expect everyone to have the average characteristics of the whole population except where individual evidence is available. Notice that generalizing over a larger group like this should increase the misrepresentation of people, and thus their inappropriate treatment. Recall that that was the original problem with discrimination.
If the parents mentioned earlier were undiscriminating they would be much more trusting of middle aged men, but they would also be less trusting of other demographics such as teenage girls. All evidence they had ever got of any group or type of person being untrustworthy would be interpreted only as weaker evidence that people are untrustworthy. This would reduce the expected trustworthiness of their best candidate, so more often they would not find it worth going out in the first place. Now the man still misses out on the position, but so does the competing teenage girl plus the parents don’t get to go out. Broadening group generalizations to the extreme makes ‘discrimination’ worse, which makes sense when we consider that discriminating between people as much as possible (judging them on their own traits) is the best way to avoid ‘discrimination’.
It may be that something else about discrimination bothers you, for instance if you are most concerned with the equality status of competing social groups, then population level generalizations are the way to go. But if you want to stop discrimination because it causes people to be treated as less than they are, then work on making it easier to discriminate between people further, rather than harder to discriminate between them at all. Help people signal their traits cheaply and efficiently distinguish between others. In the absence of perfect discrimination between individuals, the other end of the spectrum is not the next best thing, it’s the extreme of misrepresentation.