I don't assume this is easy or even possible to solve. If the model is trying to determine something like the applicant's level of adversity based on whatever info college admissions has, of course the result will have a divergence between races, and historical racism will have shaped some of your input variables (like high school) to some degree that we'll never fully understand. Using some statistical difference between races as a cost function wouldn't be satisfactory either.
All I'm saying is at the bare minimum, don't feed race as an input. It's a vague factor that the applicant has zero control over, and a model (or human reviewer) making any assumptions based on race would be inherently racist, which is unjustifiable. Another bad example is how UCSD asks how feminine (rate 0-5) the applicant considers themselves. Only use factors that have some legitimate purpose; at least the high school matters because of grade inflation, but standardized tests are actually fairer. US hiring is this way, where many companies consider it an unwanted risk to even know about protected categories (age, race, religion, marital status, etc). It's far from an airtight solution.
This isn't really about science, and the article is off-topic for HN.