Diversity in Sociotechnical Machine Learning Systems

Diversity in ML is not a single concept: The paper identifies that "diversity is not a monolithic concept" and introduces a taxonomy of different diversity types, each based on distinct philosophical rationales. This means organizations cannot simply implement "diversity" as a generic solution—they must understand which specific type of diversity addresses their particular algorithmic bias problem.

Current diversity discussions lack precision about how diversity actually works: The authors highlight a critical gap where researchers discuss diversity's benefits without clearly explaining "the precise mechanisms underpinning diversity's potential benefits," which can lead to "uninformative generalities, invalid experimental designs, and illicit interpretations of findings." This suggests many current diversity initiatives in AI may not be grounded in solid understanding of why they should work.

Different diversity rationales serve different purposes: The paper distinguishes between "epistemic and political rationales" underlying different diversity concepts, indicating that diversity efforts in ML teams or algorithms may need to serve different goals—some aimed at improving decision-making quality, others at ensuring fair representation or legitimacy in the process.

Bridging philosophy and practice is essential: By drawing on research from "philosophy, psychology, and social and organizational sciences," the paper suggests that responsible AI practitioners need to engage with deeper theoretical foundations rather than treating diversity as a surface-level fix, to understand when and why diversity interventions will actually reduce algorithmic bias.