{"id":838,"date":"2020-08-10T12:13:44","date_gmt":"2020-08-10T17:13:44","guid":{"rendered":"http:\/\/capertongillett.com\/blog\/?p=838"},"modified":"2020-08-10T12:13:44","modified_gmt":"2020-08-10T17:13:44","slug":"ai-is-racist-and-sexist-because-we-were-first","status":"publish","type":"post","link":"https:\/\/capertongillett.com\/blog\/ai-is-racist-and-sexist-because-we-were-first\/","title":{"rendered":"AI is racist and sexist because we were first"},"content":{"rendered":"\n

Artificial intelligence is a handy tool. It\u2019s also somewhat misleadingly named tool \u2014 as much emphasis as we tend to put on the \u201cintelligence\u201d part of it, it\u2019s the \u201cartificial\u201d part that could benefit from more emphasis. So I\u2019m going to talk about AI, and as befits such a thoroughly modern and high-tech subject, I\u2019m going to start by talking about this chick named Shirley circa 1950.<\/p>\n\n\n\n

Shirley you jest<\/h2>\n\n\n\n
\"\"
The actual Shirley.<\/figcaption><\/figure><\/div>\n\n\n\n

Back in the mid-1950s, when commercial photo processing was just becoming a thing and Kodak was sending its first color photo printers out into the world to independent photo labs, the company came up with a clever calibration kit. It sent out color prints and unexposed negatives, and when you processed the negatives such that they matched the prints, congratulations! Your printer was calibrated properly.<\/p>\n\n\n\n

Originally, all the prints were photos of Kodak model Shirley Page, who would have to sit for literally hundreds and hundreds of photos. Over time, of course, other models were used as well, all of them smiling, bright-eyed, brunette, white, and named Shirley. (Okay, not that last one. But the prints were still called \u201cShirley cards<\/a>.\u201d) The result? Decades and decades of perfectly-exposed white people and Black people who looked like they were photographed by Annie Leibovitz<\/a> (YEAH, I SAID IT). Kodak didn\u2019t release its first non-white Shirley card until the 1970s (and even then, it was the complaints of furniture companies and chocolate makers<\/a>, not BIPOC camera owners, that led to improvements in products and processing that resulted in film that worked for darker skin tones).<\/p>\n\n\n\n

There\u2019s no such thing as a racist photo printer. It\u2019s a machine. They have no motivation. (A photographer who never bothered to learn how to light dark skin? That\u2019s another story.) But a printer calibrated based on a flattering photo of a white model, shot by a white photographer, for use with a product that was, at the time, owned mostly by white people, was inevitably going to produce poor results for anyone who didn\u2019t look like Shirley.<\/p>\n\n\n\n

Which brings us, naturally, to Google Cloud Vision seeing bearded women in 2020.<\/p>\n\n\n\n

Bearded ladies<\/h2>\n\n\n\n

This week, marketing agency Wunderman Thompson\u2019s data group released a study<\/a> looking into how different visual AI systems handle images of men and women wearing PPE. They took 265 images each of men and women with an assortment of settings, mask types, and photo quality levels and ran them against Google Cloud Vision, IBM\u2019s Watson, and Microsoft Azure\u2019s Computer Vision to see what the systems thought they were seeing.<\/p>\n\n\n\n

All of the systems struggled to spot masks on the faces of men or women, although they were twice as good at identifying male mask-wearers than female. The AIs had some interesting thoughts about what they were seeing.<\/p>\n\n\n\n