🚨 "Fair Enough AI," by Tal Zarsky & @JaneYakowitz, discusses the lack of concrete fairness standards and the inevitable tradeoffs in fairness decisions, and it's a MUST-READ for everyone in AI. It's full of 🌶spicy statements🌶: "Given the cross-cutting goals and societal aspirations that affect how decision-making will be perceived, defining and creating a “fair” algorithm is primarily a policy task rather than a matter of technology or pure logic. This fact has been absorbed in the legal scholarship for some time. The trouble is, recent AI regulatory frameworks have demonstrated an unwillingness to state which types of unfairness will be tolerated in order to avoid other forms of unfairness. Implementing one measure to promote fairness might at time generate or exacerbate fairness on another dimension. 🌶 We suspect that vagueness and abdication of decision-making will be a feature of the AI public policy debates for the foreseeable future. 🌶 Setting priorities not only raises disagreements between regulators, it causes a good deal of heartburn for each individual lawmaker, too, who will have to answer to media inquiries, firms, and voters who come armed with examples of bias, opaqueness, inaccuracy, and privacy intrusions which will follow, no matter what option she chooses. 🌶The public is not prepared for a frank admission that it is acceptable for a large AI company to decide, in advance, that it is ok to implement an algorithm that will be wrong more often for one group than another. 🌶 Nor is it prepared to hear that the same company decided in advance to reduce accuracy for everybody in order to relieve some forms of bias (but not all)" - "🌶 Some charges of unfairness are more valid than others. An accusation that an algorithm is inaccurate, biased, overly opaque, or too gamable will be valid if the faults are unnecessary—that is, if they are known or reasonably discoverable and can be corrected without significantly degrading other forms of fairness. Thus, while we have emphasized that ethical tradeoffs must be made during AI design, 🌶 that is only true for applications and designs that have already made every Pareto-efficient improvement. If an AI application needlessly compromises accuracy, bias, or some other aspect of fairness, it deserves criticism. Any time a company can make improvements for minimal costs along the other dimensions of fairness, they should. The criticisms that worry us are those that are made without any attempt to assess whether the perceived problem is easy to fix (without tradeoffs) or is difficult, requiring compromise between values." 👉 Read the full paper below. 👉 To stay informed of the latest AI governance discussions, including 🌶 spicy research papers, join 38,300+ people who subscribe to my AI governance newsletter (below). If you have curious friends in the field, tell them to subscribe!