A toolkit of dilemmas: Beyond debiasing and fairness formulas for responsible AI/ML

One-size-fits-all fairness approaches are insufficient: The paper argues that current fair AI practices often rely on formulaic bias elimination and ethics operationalization that lack sensitivity to context and complex social phenomena, suggesting that responsible AI requires more nuanced, situation-specific approaches rather than universal technical solutions.

Data availability shapes what problems we try to solve: The "opportunity dilemma" highlights a critical chicken-and-egg problem—organizations often define AI problems based on what data they have available rather than identifying genuine problems first and then finding appropriate data, potentially leading to solutions in search of problems rather than problems driving solutions.

Scalability and accuracy come at the cost of local relevance: There's a fundamental tension between building AI systems that work broadly across many contexts (requiring large, generalized datasets) versus building systems that work well for specific communities (requiring smaller, localized data), and the paper suggests this trade-off deserves explicit acknowledgment rather than being ignored.

Technical objectivity and social context both matter: Rather than choosing between treating AI as purely technical and objective or purely socially constructed, the paper advocates for "situated reasoning" that acknowledges both the technical realities and the social dimensions of algorithmic systems simultaneously.

Responsible AI requires ongoing creative problem-solving, not just rule-following: The toolkit approach emphasizes that practitioners should actively engage with dilemmas and tensions in responsible AI rather than applying predetermined formulas, suggesting responsible AI is an iterative practice rather than a checklist.