Good models, Bad data

Doctor : Nurse

Computer programmer : homemaker

What connection did you make between these two analogies? Even if just for a split second, these analogies strike most of the masses to assign doctor and programmer to a male role, while nurse and homemaker are assigned to a female one- unless you have surpassed all societal norms and reached a true level of #wokeness. Props to you if so. But, the sad truth is that these biases are ingrained in society, and our brains’ eagerness to draw conclusions from what is laid before us cause most of us fall to victim to these sexist classifications. And guess what- that means it’s causing artificial intelligence to as well.

Of course, the past decade or so has seen a huge surge in feminist campaigns and immense progress is being made. Alas, there still is always so much more to be done. Given in the tech industry itself there are enormous disparities between males and their female counterparts, it almost seems to follow that there will be some biases in the algorithms being created in Silicon Valley and beyond. Take, for example, how for every dollar a male founder makes in the tech industry, his female counterpart is making 30 cents (a gap much larger than the national gender wage gap). Or, how in the big-name companies, like Facebook, Apple and Google, there still is a 3:1 male to female employee ratio. And we can’t forget the classic study that found men are 70% more likely to get money from the same sales pitch that a woman makes. Surely, I could go on for days, but all of these inequalities in the world seep their way into what is imagined to be an entirely unbiased thing: MATH!!

But how could math be biased? The numbers don’t lie, right? How can artificial intelligence be biased, unless the programmers are intentionally making their creations sexist, racist or partial in some other sense? It’s called bad data. As put by the colead of the Ethical AI Team at Google, Timnit Gebru, “data is generated by society.” When society is biased, then the data is bad. The examples of disparities between men and women I gave above may give a sense into why AI is learning biases. While machine learning has revolutionized technology, these miraculous systems still need a basis from which they perform their mathematical magic. In a nutshell, these algorithms learn from initial training data, that is input by humans, and form generalizations from this data and apply it to novel, unseen inputs. So, if Amazon uses an AI system trained on 10 years’ worth of applications, a majority of which are from men, it’s going to teach itself that men are better candidates (because the majority of the hires will have been men), and more men are going to be called in for interviews. And yes, this actually happened.

So how do we stop these biases? More and more research is being dedicated into investigating how to eliminate biases from AI. Which, when you think about it, is a pretty astonishing task: it means teaching these systems to ignore the bias in the world. Can humans even do this? Can we even look past the biases so ingrained in our society? Maybe your response to the analogies at the beginning of this blog give you an answer. Vulnerable groups are in AI’s potential line of fire. But, thanks to cohorts at companies like IBM, Accenture and more, there is work being done to combat these issues. Psychologists are working with programmers to study cognitive mechanisms humans use to look past biases, so that these processes can be modeled in AI.

Women and other marginalized groups shouldn’t be afraid of AI, however. It’s a relatively new part of our modern world, and computer scientists and humanitarians alike are working together to make sure that it’s helping society more than it’s harming it. In fact, many groups are making headway with eliminating biases from math in AI with, in fact, more math. Algorithms are being constructed to counteract the generalizations that other-wise impartial systems may learn from biased data. There is light at the end of the tunnel! And what we can do is not be afraid of the advancement of our technology, and help contribute to the “good” data! After all, no matter how perfect an equation is, an imperfect world will still push it to wrong conclusions.

-Daniela Torres, Sophomore Student at Johns Hopkins University