Photo credit: www.technologyreview.com
Rethinking Fairness and Bias in AI: New Directions in Research
Divya Siddarth, the founder and executive director of the Collective Intelligence Project, emphasizes that long-standing definitions of fairness and bias require reevaluation. “We have been sort of stuck with outdated notions of what fairness and bias means for a long time,” she observes. “We have to be aware of differences, even if that becomes somewhat uncomfortable.”
The recent study led by Wang and her team represents progress in this area. Miranda Bogen, director of the AI Governance Lab at the Center for Democracy and Technology, notes, “AI is used in so many contexts that it needs to understand the real complexities of society, and that’s what this paper shows.” She stresses the necessity of approaching the problem with nuance rather than oversimplification, warning that “just taking a hammer to the problem is going to miss those important nuances and fall short of addressing the harms that people are worried about.”
The benchmarks suggested in the Stanford paper aim to provide clearer guidelines for assessing fairness in AI models. However, actual improvements to these models may call for varied strategies. One potential method is to enhance the diversity of datasets used in training AI systems. Although important, gathering such data can be expensive and labor-intensive. Siddarth highlights the value of community feedback, stating, “It is really fantastic for people to contribute to more interesting and diverse datasets. Feedback from users indicating their sense of representation, or lack thereof, can be critical for refining model responses.”
In addition, the field is exploring mechanistic interpretability, where researchers analyze the inner workings of AI models. “People have looked at identifying certain neurons that are responsible for bias and then zeroing them out,” explains Augenstein. This approach underscores the potential for targeted modifications within the AI’s architecture to reduce bias at a granular level.
However, some experts, such as Sandra Wachter, a professor at the University of Oxford, express skepticism about the notion of achieving fairness without human oversight. “The idea that tech can be fair by itself is a fairy tale,” she argues. “An algorithmic system will never be able, nor should it be able, to make ethical assessments.” Wachter emphasizes that the law is a dynamic entity reflecting current ethical beliefs, which evolve alongside society.
Determining when AI models should differentiate among various groups can prompt significant debate. Cultural differences often lead to conflicting values, complicating the question of what ideals an AI should embody. Siddarth suggests “a sort of a federated model,” akin to global human rights frameworks, where distinct communities or nations develop their own specific models tailored to their values.
Tackling bias within AI will undoubtedly be a complex endeavor, regardless of the methodology adopted. However, Wang and her colleagues believe that equipping researchers, ethicists, and developers with improved frameworks is critical. “Existing fairness benchmarks are extremely useful, but we shouldn’t blindly optimize for them,” Wang asserts. “The biggest takeaway is that we need to move beyond one-size-fits-all definitions and think about how we can have these models incorporate context more.”
Source
www.technologyreview.com