Thank you very much for sharing this perspective. I think it probably does a good job standing in for how a lot of people think about these things. This is a difficult area of discourse, and I know people feel strongly about it. Obviously, the lecture you linked to makes many more distinctions. I'm going to try to simplify one of the distinctions to make it easier to talk about: that there is a tension between "procedurally fair" outcomes and "representatively fair" outcomes.
In my opinion, I think this idea is wrong, at least in the context of designing AI within contemporary American society. To frame "procedurally fair" and "representatively fair" as opposing value systems is a misunderstanding of the best arguments for a "representatively fair."
Let's use a common analogy: a race. This distinction imagines a situation where one race puts a bunch of runners at the starting line, tells them to run, and declares a winner based on who crosses the finish line first. That race is "procedurally fair" because all the participants raced under the same conditions. Then, there's a "representationally fair" race, where runners from "protected groups" are allowed to start the race from the middle of the track.
I think there are two ways that this isn't right that come from a common misperception, that the only relevant time to think about in the context of these decisions is the present. (I also want to mention that there's a lot of heavy-handed language use in this presentation that reveals the author's perspective. Algorithms tilted against black people "reveal" disparities,
First, it's important to have a more accurate perspective of the past. In most cases where "representationally fair"-type solutions are used, there is a historic reason why the "protected group" can't run as fast as the other group. If people in Hyderabad are more likely to defraud a micro-lender, that's not a spontaneous result of inherent differences between Hyderabad-type humans and other humans. A quick google search tells me that as of 2017 Hyderabad has the second largest poor populace in India. So, in the race to a successful economic outcome from birth, the so-called "procedurally fair" solution actually means taking the Hyderabad runners as infants very far back from the starting line. All else being procedurally equal, when the starting gun of "applying for a micro-loan" goes off, the Mumbaikars and Delhiites are already far ahead of the Hyderabadis. Once we have a more accurate perspective that includes the past, in order to make the race procedurally fair we have to move everyone into place at a common starting line, which will inevitably mean helping the Hyderabadis forward.
Second, it's important to acknowledge the effect of biased algorithms on the future. Not only will the unfair, pseudo-"procedurally fair" approach unjustly disadvantage some runners in this race, future races are calibrated according to achievement in past races. If you win one race, you get a head start in the next one. Micro-lenders who refuse to lend to Hyderabadis will exacerbate the relative poverty situation that the algorithm is picking up on. It's easier to pretend to yourself that you're making an algorithm that peeks into the world, makes an objective judgment, and then pops back out of the world. But in fact, people designing algorithms have a responsibility for the outcomes of their algorithms. If an unfair situation exists (for example, that just by being born in Hyderabad and not Delhi, any given person will start life with less economic power), your algorithm's consequences will either be helping to fix that unfair situation, or making it worse.
So, if your algorithm punishes Hyderabadi applicants for being poor, it is both unfair in the simplest sense once you account for where the applicants started and will increase the unfairness of any future round of applications. Put another way, what we're talking about at a high level is values. In one libertarian version of society, the purpose of the algorithm is to maximize profit for Simpl. That society values maximizing profit, and enshrines "shareholder value" as the centerpiece of its ethics. In AOC's socialist version of society, justice, equality, and eliminating poverty are valued highest. This is the broader point she's making about algorithms, that they reflect the values of the people who make them. Given the power that these algorithms have, in our example to lift people out of poverty or to deepen economic inequality, society and not just technologists (not the most diverse group in countless ways) should lead the decisions about what should be valued.
We're seeing more and more the consequences of a techno-libertarian approach, and I'm curious what we'll see as consequences if a more socialist approach wins for a while. I value fairness over shareholder value, so the prospect excites me. But I'm sure there will be lots of unintended consequences in such a system too (the over-cited "asians applying to elite colleges" example being a good case of an outcome that's not easily justified on its own. I think elite college admissions are more fundamentally broken, but that's another thread!)
