In December, a group called Black in AI plans to host an afternoon workshop to highlight AI research by black computer scientists at the Conference on Neural Information Processing Systems, one of the top gatherings for scientists working on AI. While the organisers invited people of all races to attend the workshop, they said only black researchers would be allowed to present papers. As news of the event spread on social media, it sparked a backlash from some coders and academics who questioned why an event focusing solely on research by black scientists was necessary. The debate echoes the controversy in August that followed Google employee James Damore’s circulation of a manifesto that, among other points, accused the company of overzealously promoting diversity at the expense of technical ability in hiring and promotions.
Questions about diversity in the AI and machine-learning fields focus on inclusion and employment – but the issue has even more far-reaching consequences. Software algorithms only behave effectively and fairly if they have a diverse set of data to train them. The concern is that as these automated systems are used in increasingly critical decisions, such as who gets a bank loan or is granted parole, a lack of diversity among the scientists behind them will produce biased results.
Already there have been a number of high-profile incidents in which AI programmes exhibited racial bias. In 2016, a computer programme designed to judge human beauty using supposedly objective features in photographs – like facial symmetry and wrinkles – was found to heavily favour people with light skin. That same year, after Microsoft Corp released a Twitter chatbot called Tay, users easily goaded it into using racist language and expressing neo-Nazi views. An algorithm that was supposed to provide parole boards with objective advice on whether a prisoner was likely to commit another crime if released was found to incorrectly judge black defendants as “high risk” twice as often as white defendants.
There are concerns about hidden bias within big data sets – on everything from mortgage lending and insurance underwriting to the clinical trials of new drugs – that are increasingly being used to train machine- learning algorithms.
Timnit Gebru, who researches artificial intelligence for Microsoft and is one of the few black female computer scientists specialising in the field in the US, co-founded Black in AI. She has been vocal in trying to call attention to the problem of bias in the data that machine-learning algorithms use, and has said the lack of diversity in the field means that those building AI systems are less aware of these issues than they should be.
Black in AI seeks to foster collaboration among researchers and increase the presence of black people in the field, according to its website.
The debate over the Black in AI workshop began shortly after Ian Goodfellow, a research scientist at Google Brain, the artificial intelligence lab that is part of Alphabet Inc, promoted it on his own Twitter feed on October 11. “Why is this necessary? If the black people projected to attend wanted to present their work why can’t they do so at an event for all races?” a user with the Twitter handle @typeload wrote in response to Goodfellow’s post. Others soon chimed in too. “This actually promotes segregation of AI (and society in general): blacks going to events for blacks, women – for women, e.t.c.,” tweeted Timofey Yarimov, a data scientist who works for SKB Kontur, a Russian firm that makes business and accounting software. “I don’t see this necessary, the benefits will be much less than the drawback,” tweeted Ahmed Adly, whose LinkedIn profile lists his title as “head of mobility” for Qatar-based IT services firm called Malomatia.
Yarimov, when contacted by Bloomberg, declined to answer further questions about his view.
Adly said in an email to Bloomberg that he agreed there was discrimination in the field of AI against women and non-white men, but was unsure what the true intent of the Black in AI event was. “If the event is part of an effort to fight the discrimination that we know exists then that’s a good thing,” he wrote. But if it was designed “to use the hardship of a specific category just to attract attendees, then it’s not good” and would only undermine the cause of combating discrimination. He also said that while he led his company’s efforts to look at possible uses of AI, he was not planning to attend the NIPS conference this year.
