The four-fifths rule is not disparate impact: a woeful tale of epistemic trespassing in algorithmic fairness

The "four-fifths rule" in AI fairness software is not equivalent to the legal concept it claims to represent. Computer scientists have abstracted a specific part of U.S. disparate impact discrimination law into an AI metric, but this abstraction omits crucial legal context and creates new ethical problems that don't exist in the original law.

Using this simplified metric can introduce new forms of harm. The paper warns that the "disparate impact metric" used in AI fairness toolkits changes the ethical meaning of the original legal rule and creates "new potentials for ethical harms" by removing important legal nuances about how discrimination should be evaluated.

Popular AI fairness software is spreading this flawed abstraction. The authors found that the four-fifths rule has been "codified into popular AI fairness software toolkits," meaning many organizations may be implementing a misrepresented version of anti-discrimination law without realizing it.

Computer scientists need to be more careful about borrowing concepts from other fields. The paper demonstrates a problem called "epistemic trespassing"—when experts from one field (computer science) apply concepts from another field (law) without fully understanding their original context, risking serious unintended consequences in real-world applications.

Interdisciplinary AI ethics requires genuine collaboration, not just translation. The authors call for computer scientists to "self-critically re-evaluate" their abstractions, particularly when working on AI ethics, suggesting that responsible AI development requires deeper engagement with legal and social expertise rather than simplified technical shortcuts.