“We’ve got been form of stuck with outdated notions of what fairness and bias means for a very long time,” says Divya Siddarth, founder and executive director of the Collective Intelligence Project, who didn’t work on the brand new benchmarks. “We’ve got to pay attention to differences, even when that becomes somewhat uncomfortable.”
The work by Wang and her colleagues is a step in that direction. “AI is utilized in so many contexts that it needs to grasp the true complexities of society, and that’s what this paper shows,” says Miranda Bogen, director of the AI Governance Lab on the Center for Democracy and Technology, who wasn’t a part of the research team. “Just taking a hammer to the issue goes to miss those vital nuances and [fall short of] addressing the harms that individuals are anxious about.”
Benchmarks just like the ones proposed within the Stanford paper could help teams higher judge fairness in AI models—but actually fixing those models could take another techniques. One could also be to take a position in additional diverse data sets, though developing them could be costly and time-consuming. “It is absolutely implausible for people to contribute to more interesting and diverse data sets,” says Siddarth. Feedback from people saying “Hey, I don’t feel represented by this. This was a very weird response,” as she puts it, could be used to coach and improve later versions of models.
One other exciting avenue to pursue is mechanistic interpretability, or studying the inner workings of an AI model. “People have checked out identifying certain neurons which can be accountable for bias after which zeroing them out,” says Augenstein. (“Neurons” on this case is the term researchers use to explain small parts of the AI model’s “brain.”)
One other camp of computer scientists, though, believes that AI can never really be fair or unbiased and not using a human within the loop. “The concept tech could be fair by itself is a fairy tale. An algorithmic system won’t ever find a way, nor should it find a way, to make ethical assessments within the questions of ‘Is that this a desirable case of discrimination?’” says Sandra Wachter, a professor on the University of Oxford, who was not a part of the research. “Law is a living system, reflecting what we currently consider is moral, and that ought to move with us.”
Deciding when a model should or shouldn’t account for differences between groups can quickly get divisive, nevertheless. Since different cultures have different and even conflicting values, it’s hard to know exactly which values an AI model should reflect. One proposed solution is “a form of a federated model, something like what we already do for human rights,” says Siddarth—that’s, a system where every country or group has its own sovereign model.
Addressing bias in AI goes to be complicated, irrespective of which approach people take. But giving researchers, ethicists, and developers a greater origin seems worthwhile, especially to Wang and her colleagues. “Existing fairness benchmarks are extremely useful, but we shouldn’t blindly optimize for them,” she says. “The most important takeaway is that we want to maneuver beyond one-size-fits-all definitions and take into consideration how we will have these models incorporate context more.”
