Search engines can reveal some unattractive biases when you squint at them closely. Check out what Google autofills, or displays in its top results, and you’ll see just how entrenched our collective prejudices are re: stuff like “feminine” work or black criminality.
The paper — which has been submitted to the International Conference on Social Informatics but has not yet been published — looked at how Google and Bing represent female beauty in their image search results, particularly when it comes to different age and racial groups. To do that, a graduate student and two professors, Virgilio Almeida and Wagner Meira Jr., scraped the top 50 images for “beautiful woman” and “ugly woman” across dozens of international versions of Google and Bing. They then passed those 2,000-plus images through a program called Face++, which estimates subject age, race and gender with 90 percent accuracy.
Sadly, the race and age breakdown of the “beautiful” pictures versus the “ugly” ones are probably what you’d expect. For almost every country analyzed, white women appear more in the “beautiful” results, and black and Asian women appear in the “ugly” ones.
Blackness is considered less attractive in 86 percent of the countries surveyed on Google, including countries such as Nigeria, Angola and Brazil, where there’s a predominance of people with black or brown skin.
Likewise, beauty is associated almost exclusively with youth — extreme youth, in some countries. In Japan and Malaysia, for instance, queries for “beautiful woman” don’t turn up ladies much older than 23.
In the United States, searches for “beautiful” women result in pictures that are 80 percent white, and roughly between the ages of 19 and 28. Searches for “ugly” women are roughly 60 percent white and 30 percent black, and fall into the 30 to 50 age range.
This sort of bias has been observed on search engines before, of course — although, in a typical chicken/egg quandary, it remains unclear whether the results shape society, or society shapes the search results. A 2013 paper by Harvard professor Latanya Sweeney found that searches for black names surface more ads related to criminal record history, whether or not the person in question was associated with any crimes. Likewise, a 2015 study from researchers at the University of Washington and the University of Maryland found that Google almost exclusively displays pictures of men for queries like “construction worker.”
In 2013, Paul Baker and Amanda Potts, linguistic researchers from Lancaster University in Great Britain, conducted a survey of Google Instant’s autofill results — you know, the zany suggestions Google comes up with while you’re still typing your search term — and observed that racist and homophobic language frequently made it into Google’s suggested answers. They don’t fault Google, though — rather, they conclude that the search engine started predicting racist, homophobic stuff because its users kept entering it. Google is, in other words, rather like our ex-friend Tay: a product of her sorry circumstances.
The authors of this latest paper aren’t quite so comfortable absolving Google and Bing of blame. Yes, acknowledged Almeida, one of the co-authors of the paper and a current visiting professor at Harvard: Pre-existing social biases absolutely shape image search results — someone had to upload and tag and post these photos in the first place, after all. But “the way search engines index and rank images” could also contribute to the creation, or at least the enforcement, of stereotypes, he explained by email.
“We do not have [enough] information about the techniques used by search engines to rank images and photos,” he said.
It’s unlikely that researchers such as Almeida and his co-authors will ever get access to that type of information: Google and Bing guard their ranking algorithms like state secrets. But Almeida says another one of his and Meira’s graduate students is working on “techniques to increase [the] transparency of platforms,” and that both computer scientists remain interested in auditing, and even reverse-engineering, them.
Until they have more complete information, however, the researchers are hopeful that companies such as Google and Microsoft will begin re-evaluating their own algorithms.
“Given the importance of search engines as source[s] of information,” they write, “we suggest that they analyze the … prominent presence of negative stereotypes and find algorithmic ways to minimize the problem.”
Liked that? Try these! Which are, er — not awesome.