How to create unbiased algorithms in a biased society

Artificial intelligence systems are meant to transcend some of the imperfections of the human mind, being grounded in mathematics and operating on data rather than emotion or subjective perception. But as more algorithms are weaved into daily life, the limits of their objectivity are being revealed.

Take Google’s speech recognition algorithm. It’s a handy tool that allows you to perform an internet search by asking Google what you want to know. The algorithm is trained by listening to audio data and learning patterns. But research has shown that it has a gender bias—it performs better with male voices than female voices. This isn’t because Google’s algorithm is sexist, but because it was trained on unbalanced data with disproportionately more male voices, said Tina Eliassi-Rad, associate professor of computer science at Northeastern and a core faculty member in the university’s Network Science Institute.

Eliassi-Rad, whose research is rooted in data mining and machine learning, is interested in exploring how to create fair artificial intelligence systems in an unjust world. She recently gave a public lecture on the topic, titled “Just Machine Learning,” hosted by Harvard University’s Edmond J. Safra Center for Ethics.

“Algorithms can amplify the biases that we have in society,” Eliassi-Rad explained, noting several other examples that show how algorithms operating on biased data can lead to racist or sexist outcomes. A Nikon camera with built-in facial recognition software asks the user, “Did someone blink?” when the subject of the photo is Asian. In 2015, Google’s facial recognition algorithm identified two African American people as “gorillas.”

As with the speech gender bias problem, these inaccuracies stem from a lack of sufficiently diverse data used in training the algorithm. “These embarrassing incidents could’ve been avoided if the machine learning algorithm was given more balanced training data,” Eliassi-Rad said.

However, there are other algorithms with higher stakes and more malevolent consequences. For example, a ProPublica investigation last year found that a private software used to predict future criminals was biased against black people.

“The software provided the judge with an individual’s risk assessment score. Unfortunately, the scores had disproportionately high false positive rates for blacks and disproportionately high false negative rates for whites,” Eliassi-Rad said. “The ProPublica investigation details how the scores affected the decisions of judges—in one case overturning a plea deal. This study had a big impact on research in fairness and machine learning, in part because the dataset was released by ProPublica.”

Part of the problem in creating fair algorithms is the concept of fairness itself. What’s considered fair and precise in the field of computer science may not translate well to justice in the real world. One way to address this problem is by putting computer scientists into conversation with ethicists, philosophers, and others from fields that have historically examined justice and fairness.

“We need to work with other disciplines that have spent decades studying what is fair and what is just to come up with different definitions of fairness,” Eliassi-Rad said.