Artificial Intelligence—With Very Real Biases
Artificial Intelligence—With Very Real Biases
According to Microsoft researcher Kate Crawford, digital brains can be just as error-prone and biased as ours
By Kate Crawford Oct. 17, 2017 11:05 a.m. ET
What do you imagine when someone mentions artificial intelligence? Perhaps it’s something drawn from science-fiction films: Hal’s glowing eye, a shape-shifting terminator or the sound of Samantha’s all-knowing voice in the movie “Her.”
As someone who researches the social implications of AI, I tend to think of something far more banal: a municipal water system, part of the substrate of our everyday lives. We expect these systems to work—to quench our thirst, water our plants and bathe our children. And we assume that the water flowing into our homes and offices is safe. Only when disaster strikes—as it did in Flint, Mich.—do we realize the critical importance of safe and reliable infrastructure.
Artificial intelligence is quickly becoming part of the information infrastructure we rely on every day. Early-stage AI technologies are filtering into everything from driving directions to job and loan applications. But unlike our water systems, there are no established methods to test AI for safety, fairness or effectiveness. Error-prone or biased artificial-intelligence systems have the potential to taint our social ecosystem in ways that are initially hard to detect, harmful in the long term and expensive—or even impossible—to reverse. And unlike public infrastructure, AI systems are largely developed by private companies and governed by proprietary, black-box algorithms.
A good example is today’s workplace, where hundreds of new AI technologies are already influencing hiring processes, often without proper testing or notice to candidates. New AI recruitment companies offer to analyze video interviews of job candidates so that employers can “compare” an applicant’s facial movements, vocabulary and body language with the expressions of their best employees. But with this technology comes the risk of invisibly embedding bias into the hiring system by choosing new hires simply because they mirror the old ones. What if Uber, with its history of poorly behaved executives, used a system like this? And attempting to replicate the perfect employee is an outdated model of management science: Recent studies have shown that monocultures are bad for business and that diverse workplaces outperform more homogenous ones.
New systems are also being advertised that use AI to analyze young job applicants’ social media for signs of “excessive drinking” that could affect workplace performance. This is completely unscientific correlation thinking, which stigmatizes particular types of self-expression without any evidence that it detects real problems. Even worse, it normalizes the surveillance of job applicants without their knowledge before they get in the door.
These systems “learn” from social data that reflects human history, with all its biases and prejudices intact. Algorithms can unintentionally boost those biases, as many computer scientists have shown. Last year, a ProPublica expose on “Machine Bias” showed how algorithmic risk-assessment systems are spreading bias within our criminal-justice system. So-called predictive policing systems are suffering from a lack of strong predeployment bias testing and monitoring. As one RAND study showed, Chicago’s algorithmic “heat list” system for identifying at-risk individuals failed to significantly reduce violent crime and also increased police harassment complaints by the very populations it was meant to protect. We have a long way to go before these systems can come close to the nuance of human decision making and even further until they can offer real accountability.
Artificial intelligence is still in its early adolescence, flush with new capacities but still very primitive in its understanding of the world. Today’s AI is extraordinarily powerful when it comes to detecting patterns but lacks social and contextual awareness. It’s a minor issue when it comes to targeted Instagram advertising but a far more serious one if AI is deciding who gets a job, what political news you read or who gets out of jail.
AI companies are now targeting everything from criminal justice to health care. But we need much more research about how these systems work before we unleash them on our most sensitive social institutions. To this end, I’ve been working with both academic and tech industry colleagues to launch The AI Now Institute, based at New York University. It’s a multidisciplinary center that brings together social scientists, computer scientists, lawyers, economists, and engineers to study the complex social implications of these technologies.
As the organizational theorist Peter Drucker once wrote, we can’t manage what we can’t measure. As AI becomes the new infrastructure, flowing invisibly through our daily lives like the water in our faucets, we must understand its short- and long-term effects and know that it is safe for all to use. This is a critical moment for positive interventions, which will require new tests and methodologies drawn from diverse disciplines to help us understand AI in the context of complex social systems. Only by developing a deeper understanding of AI systems as they act in the world can we ensure that this new infrastructure never turns toxic.
—Crawford is a distinguished research professor at NYU and a principal researcher at Microsoft