In fact, what we’ve described here is actually a best-case scenario where fairness can be enforced by making simple changes that affect individual team performance. In practice, fairness algorithms can be more radical and unpredictable. Investigation found that, on average, most computer vision algorithms improve fairness by harming all groups—for example, by reducing recall and accuracy. Unlike in our hypothesis, where we reduced damage to a group, it is possible that the level reduction could make everyone worse off directly.
Leveling run down against the goals of algorithmic equity and broader goals of equality in society: to improve outcomes for historically marginalized or marginalized groups. Lowering performance on high-achieving groups does not naturally benefit lower-performing groups. Furthermore, the leveling can direct harm to historically disadvantaged groups. Choosing to remove a benefit instead of sharing it with others shows a lack of interest, solidarity, and a willingness to take the opportunity to actually fix the problem. It stigmatizes historically disadvantaged groups and reinforces the social segregation and inequality that has led to the problem in the first place.
As we build AI systems to make decisions about people’s lives, our design decisions encode latent value judgments about what should be prioritized. Downgrading is the consequence of choosing to only measure and address equity based on disparities across groups, while ignoring the utility, welfare, privileges, and other goods that are central to questions about equality in the real world. That is not the inevitable fate of algorithmic fairness; rather, it is the result of following the path of least mathematical resistance, and not for any overarching social, legal, or ethical reason.
To move forward, we have three options:
• We can continue to implement biased systems that benefit only a privileged section of the population while seriously harming others.
• We can continue to define equity in formal mathematical terms and implement AI that is less accurate for all groups and positively harmful for some.
• We can act and achieve justice through “leveling up”.
We believe that leveling up is the only morally, ethically and legally acceptable path. The challenge for the future of fairness in AI is to create systems that are fundamentally fair, not just procedurally fair through decentralization. Leveling up is a more complex challenge: It needs to be combined with aggressive steps to find the actual cause of deviations in the AI system. Technical solutions are often just an initial support solution to deal with a broken system. Improve access to healthcare, manage more diverse datasets, and develop tools that specifically target issues faced by previously disadvantaged communities can help make substantive equity a reality.
This is a much more complex challenge than just tweaking a system to make two numbers equal across groups. It may require not only significant technological and methodological innovations, including redesigning AI systems from the ground up, but also significant societal changes in areas such as access and spending of care. health.
While it may be difficult, this refocusing on “fair AI” is essential. AI systems make life-changing decisions. The choices about how and for whom to be fair are too important to treat equity as a simple problem to be solved. This is the status quo that has led to fair methods achieving equality through downgrading. So far, we have produced methods that are mathematically fair, but cannot and do not clearly benefit disadvantaged groups.
This is not enough. Current tools are seen as a solution to algorithmic fairness, but so far they have not delivered on their promise. Their shady moral effects make them less likely to be used and can slow down real solutions to these problems. What we need are systems that are fair through leveling up, helping underperforming teams without arbitrarily harming others. This is the challenge that we must now address. We need AI that is intrinsically fair, not just mathematical.
