Most enterprises deploying AI tools today are aware of the potential risks of bias. An algorithm that discriminates against certain classes of people is plainly unacceptable, and most AI practitioners are at least somewhat attuned to this risk.
But a deeper problem lurks inside AI systems that could undermine enterprises' good intentions. It's a problem that, by definition, few are even aware of: unconscious algorithmic bias.
"Bias is insidious because if you have biased data, it doesn't matter how much you have, it's still biased. And those kinds of errors aren't going to be washed out in the quantity of data," said Cheryl Martin, chief data scientist at Alegion, an AI training platform.
Several high-profile and obvious examples of biased AI have put the issue of algorithmic bias on many AI practitioners' agendas. Many are familiar with Microsoft's chatbot Tay, which quickly devolved into a racist and misogynistic troll, or the Google image classification service that classed images of African-Americans as images of gorillas.
Enterprises now know to scratch AI tools that deliver these kinds of obviously biased results. But relatively few are thinking about bias in a more systematic way.
In a recent survey conducted by O'Reilly Media, just 40% of respondents said they check for fairness and bias when building models. Only 17% said they consider measures of bias and fairness when evaluating the success of their models.
Bias often starts with bad data
Without systematically attacking algorithmic bias, it can creep into models in subtle ways.
For example, Martin used the example of an image recognition tool that aims to identify the gender of people in images. Most stock libraries have higher percentages of images of women in kitchens than men, so the algorithm will be trained to expect women to be in pictures in kitchens. This isn't a problem with the algorithm. There was no malicious intent on the part of the developer. The algorithm simply encoded a pre-existing bias.
Martin said addressing this kind of unconscious algorithmic bias requires users to curate data sets before using the data in training algorithms to ensure equal representation. Engineers should also review algorithms' outputs and check for apparent bias before putting AI tools out in the world.
"To fix bias, you have to be aware that it's arising from the data, and you have to correct for the bias that's in the data," Martin said.
Build a diverse team
For Sanjay Srivastava, chief digital officer at consulting firm Genpact, building diverse teams also helps address the issue of unconscious bias. Involving people who come from different backgrounds can help a team identify potential edge use cases or gaps in training data that could cause an AI service to work better for certain populations than others.
He said the key to eliminating unconscious algorithmic bias is to get a broad view of how your tool works and what it aims to accomplish. Only then can you see the areas in which it falls short.
"You have to have people that can bring in common sense and be attuned to political issues and historical context," Srivastava said. "You have to have a diverse set of backgrounds to pick up that nuance."
Srivastava said examining every AI tool for potential unconscious bias requires a tradeoff. On the one hand, engineers are under pressure to deliver working products to business units quickly, as time to market is usually a factor. But, on the other hand, involving broad teams and making efforts to eliminate bias can be time-consuming. As the issue of unconscious bias grows in prominence, he expects more enterprises to figure out how to strike this balance.
"You'd be hard pressed to find companies that aren't working on conscious bias," he said. "The issue is unconscious bias. I think the future of AI is very bright. There will be bumps in the road, but you have to learn, and I think it's important to be vigilant."