 Hello, my name is Joy Blomwini, founder of the Algorithmic Justice League, where we focus on creating a world with more inclusive and ethical AI systems. The way we do this is by running algorithmic audits to hold companies accountable. Also a point of code telling stories that make daughters of diasporas dream and sons of privilege pause. So today it's my pleasure to share with you a spoken word poem that's also an algorithmic audit called AI Ain't I a Woman. And it's a play on Sir Jernert Ruse's 19th century speech where she was advocating for women's rights asking for her humanity to be recognized. So we're going to ask AI if it recognizes the humanity of some of the most iconic women of color. You ready? My heart smiles as I bask in their legacies, knowing their lies have altered many destinies. In her eyes I see my mother's poise. In her face I glimpse my auntie's grace. In this case of deja vu, a 19th century question comes into view. In a time when Sir Jernert Ruse asked, ain't I a woman? Today we pose this question to new powers making bets on artificial intelligence hope towers. The Amazonians peek through windows blocking deep blues as faces increment scars. Old burns, new urns, collecting data, chronicling our past, often forgetting to deal with gender, race, and class. Again I ask, ain't I a woman? Face by face the answers seem uncertain. Young and old proud icons are dismissed. Can machines ever see my queens as I view them? Can machines ever see our grandmothers as we view them? Ida B. Wells, data science pioneer, hanging facts, stacking stats on the lynching of humanity, teaching truths hidden in data, each entry and omission, a person worthy of respect. Shirley Chisholm unbought and embossed the first black congresswoman, but not the first to be misunderstood by machines well versed in data driven mistakes. Michelle Obama unabashed and unafraid to wear her crown of history, yet her crown seems a mystery to systems unsure of her hair, a wig, a buffon, a toupee, maybe not. Are there no words for our braids and our locks? The sunny skin and relaxed hair make Oprah the first lady. Even for her face well known, some algorithms fault her, echoing sentiments that strong women are men. We laugh celebrating the successes of our sisters with Serena's smiles. No label is worthy of our beauty. So there you are. And so what you see in the poem I just shared which is also an algorithmic audit is a reflection of something I call the coded gaze. Now you might have heard of the white gaze, the male gaze, the post-colonial gaze, well to this lexicon. We add the coded gaze and it is a reflection of the priorities, the preferences, and also sometimes the prejudices of those who have the power to shape technology. So this is my term for algorithmic bias that can lead to exclusionary experiences or discriminatory practices. So let me show you how I first encountered the coded gaze. I was working on a project that used computer vision, didn't work on my face until I did something. I pulled out a white mask and then I was detected. So I wanted to know what was going on and I shared this story with a large audience using the TED platform over a million views and I thought somebody might check my claims. So let me check myself. And I took my TED profile image and ran it on the computer vision systems from many leading companies and I found some companies didn't detect my face at all. But the companies that did detect my face labeled me male. I'm not male. I'm a woman phenomenally. And so I wanted to know what was going on. Then I read a report coming from Georgetown law showing that one in two adults, over 130 million people, has their face in a face recognition network that can be searched unwarranted using algorithms that haven't even been audited for accuracy. And across the pond in the UK where they actually have been checking how these systems work, the numbers don't look so good. You have false match rates over 90% more than 2,400 innocent people being mismatched. And you even had a case where two innocent women were falsely matched with men. So some of the examples that I show in AI, anti-woman or the TED profile image, they have real world consequences. And because of the real world consequences, this is why I focused my MIT research on analyzing how accurate systems were when it came to detecting the gender of a particular face. And so with the research we're doing, it's been actually covered in more than 30 countries, more than 240 articles, talking about some of the issues with facial analysis technology. So in order to assess how well these systems actually work, I ran into a problem. A problem that I call the pale male data issue. And so in machine learning, which are the techniques being used for computer vision, hence finding the pattern of a face, data is destiny. And right now, if we look at many of the training sets or even the benchmarks by which we judge progress, we find that there's an overrepresentation of men, 75% male, or this national benchmark from the US government, 80% lighter skin individuals. So pale male data sets are destined to fail the rest of the world, which is why we have to be intentional about being inclusive. So the first step was making a more inclusive data set, which we did called the pilot parliament's benchmark, which was better balanced by gender and skin type. The way we achieved better balance was by going to the UN women's website, and we got a list of the top ten nations in the world by their representation of women. Rwanda leading the way, progressive Nordic countries in there, and a few other African countries as well. We decided to focus on European countries and African countries to have a spread of skin types. So finally, with this more balanced data set, we could actually ask the question, how accurate are systems from companies like IBM, Microsoft, face plus plus leading billion dollar tech company in China used by the government when it comes to guessing the gender of a face? So what do we see? The numbers seem okay. 88 maybe get a B with IBM, 94%, Microsoft is the best case overall, and face plus plus is in the middle. Where it gets interesting is when we start to split it down. So when we evaluate the accuracy by gender, we see that all systems work better on male faces than female faces across the board. And then when we split it by skin type, again we're seeing these systems work better on lighter faces than darker faces. Then we did something that hadn't been done in the field before, which was doing an intersectional analysis barring from some of Kimberly Crenshaw's work on anti-discrimination law, which showed that if you only did single axis analysis, right, so if we only look at skin type, if we only look at gender, we're going to miss important trends. So taking inspiration from that work, we did this intersectional analysis. And this is what we found. For Microsoft, you might notice that for one group, there is flawless performance. Which group is this? The pale males for the win. And then you have not so flawless performance for other groups. So in this case, you're seeing that the darker skin females are around 80%. These were the good results. Let's look at the other companies. So now let's look at face plus plus. China has the data advantage, right? But the type of data matters. And so in this case, we're actually seeing that the better performance is on darker males marginally. Again, you have darker females with the worst performance. And now let's look at IBM. For IBM, lighter males take the lead again. Here you see that for lighter females, there's a disparity, right, between lighter males and lighter females. But lighter females actually have a better performance than darker males. And then categorically across all of these systems, darker females had the worst performance. So this is why the intersectional analysis is important, because you're not going to get the full spectrum of what's going on if you only do single axis analysis. Now we took it even further and we disaggregated the results of the darker female, since that was the worst performing group. And this is what we got. We got air rates as high as 47% on a task that has been reduced to a binary. Gender's more complex than this. But the systems we tested used male and female labels, which means they would have a 50-50 shot of getting it right by just guessing. So for these systems, we're paying to do an audit that actually shows is marginally better than chance. So I thought the companies might want to know what was going on with their systems and I shared the research. IBM was by far the most responsive company, got back to us the day we shared the research and in fact released a new system when we shared the research publicly. So first, we gave the research privately to all of the companies and gave them some time to respond. So here you can see that there's a marked improvement from 2017 to 2018. So for everybody who watched my TED talk and said, isn't the reason you weren't detected because of physics, your skin reflectance, contrast, et cetera. The laws of physics did not change between December 2017 when I did the study and 2018 when they launched the new results. What did change was they made it a priority and we have to ask why. So this past summer, you actually had an investigative piece that showed that IBM reportedly secretly supplied the New York police department with surveillance tools that could analyze video footage by skin type, skin color in this case and also the kind of facial hair somebody had or the clothing that they were wearing. So enabling tools for racial profiling. And then for the New York Times, I wrote an op-ed talking about other dangers, use of facial analysis technology. So you have a company called HireView, for example, that says we can use verbal and nonverbal cues according to their marketing materials and infer somebody's going to be a good employee for you. And how we do this is we train on the current top performers. Now, if the current top performers are largely homogeneous, we could have some problems. So it's not just the question of having accurate systems, right? How these systems are used is also important. And this is why we've launched something called the Safe Face Pledge. And the Safe Face Pledge is meant to prevent the lethal use of facial analysis technology, don't kill people with face recognition, very basic. And then also thinking through things like secret mass surveillance or also the use by law enforcement. So so far we have three companies that have come on board to say we're committed to the ethical and responsible development of facial analysis technology. And we also have others who are saying we'll only purchase from these companies. So if this is something that you're interested in supporting, please consider going to the Safe Face Pledge site. And if you wanna learn more about the Algorithmic Justice League, visit us at AJLUnited.org. Thank you.