Automatically Reducing AI Bias with Synthetic Data

Wouldn't this approach be much more likely to introduce bias since it uses prescriptive data?

It it is meant to produce a desired AI behavior, it could be useful, but the argument is off in my opinion.