What is disparate impact?

Disparate impact is a concept in the field of fairness in machine learning and AI ethics, which describes a situation where an algorithm or model disproportionately negatively affects a specific group of people, even if there is no explicit bias in the model (for example, based on race, gender, or age).

This is indirect discrimination: the model may use seemingly neutral features, but the outcome leads to systematic infringement of rights or opportunities for one of the groups.

Example: credit scoring, where the algorithm more frequently denies applicants from a certain ethnic group, even though ethnicity is not directly used.

Metrics are used to identify disparate impact, for example, comparing the share of positive decisions for different groups. If the difference is too large, it indicates the presence of disparate impact and the need to adjust the model to improve fairness.