Why can't AI distinguish human sex?

Scarce transgender image data makes it difficult for developers to build AI that distinguishes human sex.

Washington and more than 10 other states in the US will soon offer a third gender option "X" on driver's licenses. Airlines such as United Airlines have allowed customers to choose the pronouns "X" or "Mx". On Instagram, more than 16 million posts are transgender tagged.

Gender-specific AI works well only for people with typical male and female looks and between the ages of 25 and 35. (Image: CNN)

In the face of social change, sex-driven AI systems are raising concerns, especially among transgender communities, about the risk that the technology is being used for more sensitive purposes, such as wanted Crime or candidate screening.

According to CNN , gender-specific artificial intelligence (AI) does not exist in software. The site quotes Os Keyes, a transgender graduate student at the University of Washington, that the software only reinforces the developer 's gender - based construction system.

Many AI experts also express concern that the technology is adversely affected, due to the incorrect prediction of the gender of transgender people, as well as those without typical male or female looks. " We should consider the assumption that someone whose body, face, hair or style does not reflect their emotional state or gender , " said Mederith Whittaker, a former Google employee and a bright fellow. set up AI Now Research Institute of New York University, said.

Deficiencies of sexist AI

Most AI capable of determining the sexes of Microsoft, Amazon or IBM is now a binary classification system. The results can only be male or female, based on face recognition algorithms and other identifying features, such as makeup, glasses . Recently, Amazon has introduced the application of AI projections. guess gender, help store owners to classify and statistic the number of male and female customers via security video.

Kyle Martinowich, marketing director of start-up AI research startup Clarifai, said the majority of the company's customers are retail stores."The US government also uses this technology to gather information about the number of men and women at the airport or in and out of federal buildings , " Martinowich said.

AI misidentified Morgan Klaus Scheuerman as a male, instead of female as her true gender.(Photo: CNN).

Research results of Morgan Klaus Scheuerman, a PhD student at Colorada University, show that gender-specific AI still has many limitations. For example, the accuracy rate drops significantly when the system has to determine the gender of transgender people.

After analyzing with facial recognition systems of Microsoft, Amazon, IBM and Clarifai, the results showed that 98% of cases identify the right women, 97.6% cases identify the right men. However, only 87.5% of photos tagged with "transgender women" are predicted to be female and 70% of photos tagged with "transgender men" are identified as male.

According to Scheuerman, the research results show that developers have not used transgender images to train AI. Regardless of the developer's purpose, the lack of input training data can make the system work " impersonal ".

Gender-specific AI Amazon Rekcognition repeatedly predicts the gender of black women.(Photo: Amazon).

Typically, face recognition and gender discrimination are often used by police or in stores. The deployment process is rarely modified or disclosed publicly.

Moreover, technology is not always used for the original purpose of the developer. For example, Amazon stated in the Rekognition manual that "was not designed for human sex classification." However, Woo, the dating app in India, is still using Rekognition to check the gender of the user declared in the profile and pairing suggestions based on avatars.

Should we put our faith in sexist AI?

Technology companies are still working to improve the effectiveness of gender-specific AI.(Photo: CNN)

Despite ethical concerns, businesses still believe that accurate gender data will provide value. Finally, instead of giving up, some companies are struggling to find ways to improve AI predictability.

Limbik, a New York-based start-up, is developing video analytics technology to identify content that attracts viewers. The company has shifted from Amazon's AI to IBM's AI to determine the gender of people and the frequency of male appearances in some type of advertising.

Zach Schwitzky, CEO of Limbik, said the company had "struggled with the binary classification system". The two most common problems of sexist AI are the system that incorrectly identifies short-haired women as "men" , as well as the gender of teenagers. According to Schwitzky, AI only works best for people aged 25 to 35.

Currently, Limbik is building a new AI capable of classifying gender into three categories: men, women and other genders. Initially, the company's employees would have to manually arrange portraits from sites like Facebook, Twitter, Instagram . into the right categories to train AI.

Aaron Smith, laboratory director of the Pew Research Center, says the solution to gender identity and gender identity is the "subject of special interest" by AI researchers. However, he is not sure whether future technology can help accurately distinguish gender, and AI can still give false results if only judged by an appearance.