Bias in context: What to do when complete bias removal is not an option

It is widely recognized that machine learning algorithms may be biased in the sense that they perform worse on some demographic groups than others. This motivates algorithmic development to remove algorithmic bias, which in turn might lead to a hope—even an expectation—that algorithmic bias can be mitigated or removed (1). In this short comment, we make three points to qualify Wang et al.’s suggestion: 1) It may not be possible for algorithms to perform equally well across groups on all measures, 2) which inequalities count as morally unacceptable bias is an ethical question, and 3) the answer to the ethical question will vary across decision contexts.