Google’s Autocomplete search tool has, at various times, been found spreading Holocaust denial, championing white supremacy and suggesting that Muslims are “bad”, Jews are “evil” and that the Earth is, in fact, flat. But what about when it’s put to work on the images we want to see?
In September 2018, Google introduced a suggestion feature for image searches – essentially a version of Autocomplete for the visual world. Type ‘dogs’ into Google Images and a carousel of adorable puppers will appear immediately below the search bar, suggesting related image results for pugs, labradors, pitbulls and rottweilers. So far, so cute.
But switch your area of interest to, say, Hollywood actors, and Google’s algorithm gets a bit creepy. If you’re searching for a man, Google’s image-hunting algorithm will mostly focus on his career. If you’re searching for a woman, Google’s algorithm will focus on her body.
Do a Google Images search for Robert Downey, Jr., for example, and only four out of the 30 algorithmically-generated related search terms relate to his physical appearance: workout, body, handsome, cute. The vast majority focus on films he has starred in and actors he has appeared alongside. Do the same for Avengers co-star Scarlett Johansson and Google will recommend no fewer than 14 related visual searches based on her physical appearance, from “stomach” to “figure” and “big” to “rack”. It’s the male gaze, algorithm style.
To test the scale of the problem, we carried out Google Images searches for the top 60 female and male actors based on career box office earnings. As with Downey Jr. and Johansson, we collected any related terms Google suggested that were based on physical appearance. Across the 30 male actors, Google suggested we might like to see a total of 82 related searches for physical attributes – “body”, “hair” and “workout” being the runaway favourites.
For the 30 female actors, Google suggested a total of 176 related searches for physical attributes – from "belly button”, “age 21”, “oops”, “beach” and “bathing suit” to “measurement”, “thicc”, “thigh” and "pokie”. Across the search results we analysed for the top 60 female and male actors, Google’s algorithm suggested more than twice as many related searches for physical attributes of women as it did men. After WIRED brought these results to Google’s attention, it removed three of the search suggestions: “thicc”, “oops” and “pokie”.
It’s the visual equivalent of the YouTube Autocomplete Challenge, where celebrities are coaxed into answering the most-googled questions about them. “Is Emma Watson French?”, “Is Emma Watson married?”, “Is Emma Watson vegan?” – you get the idea. Crucially, these Autocomplete suggestions, while produced automatically, aren’t solely based on the popularity of search terms. In fact, over the years Google has tweaked and fiddled with its Autocomplete recipe – at one point, it launched an offensive against foot fetishists. But, when Google’s search algorithm is concerned solely with what we want to see, it still objectifies. Search for Emma Watson on Google Images and the algorithm will suggest, amongst other things, “full body”. Google’s related visual search terms for Cameron Diaz include 22 based on her physical appearance. Matt Damon, by comparison, has two.
“These results from image search are based on the underlying distribution of data about Scarlett Johansson and her mentions in text as well as image filenames and titles,” says Aylin Caliskan, assistant professor at the department of computer science at George Washington University in Washington, D.C. She points to similar, more well-known examples of biased and discriminatory Google Images search results for doctor (almost all white men); CEO (similarly male and caucasian) and unprofessional hair (mostly black women).
Such search results are the messy output of reality being pushed through a feedback loop, where real-world biases are rattled around Google’s algorithm and spat back out. For female actors, leering press coverage of “plunging necklines” and “beach-ready” bodies has, in part, informed what Google thinks is most important and relevant. But that’s not the whole story.
“Addressing these could be a long term effort that requires representing all groups of people with the same number of data points, removing historical injustices from the datasets, increasing diversity in AI developers, and democratization of AI,” says Caliskan. Case in point: 77 per cent of Google’s global technology workforce is male. In senior leadership roles, Google is 73 per cent male. Yes, the company’s search algorithm is a reflection of what we, the people who use the internet, search for. But it’s also shaped by the constraints, and priorities, hard-coded into it by Google engineers. And right now the majority of those engineers and key decision makers are men.
Read more: How to delete your Google search history and stop tracking
“We have policies against predictions that associate potentially disparaging or sensitive terms with named individuals, and when we find these predictions we work to quickly to remove them,” a Google spokesperson says. “We feel a deep responsibility not to reinforce potentially harmful perceptions in our products, and to that end we continue to invest in efforts like machine learning fairness.”
This isn’t a problem unique to Google – the issue of algorithmic bias will likely be one of the defining technological issues of the modern era. In a 2017 study, researchers at the University of Virginia noticed how data used by image-recognition software displayed gender bias. Images of shopping and cleaning, for example, were linked to women. The machine-learning software trained on these datasets didn’t just repeat these biases, it made them worse. So while the data was biased against women, the algorithmic interpretation of it was even more biased. Bad data in, even worse data out.
In this respect, it’s wrong to think of Google’s search results as a pure reflection of our own biases. The company can, and does, tweak its algorithm to hide objectionable results. Before Google tackled widespread problems with text-based Autocomplete suggestions, typing in “did the hol” returned a suggestion for “did the Holocaust happen” and then directed people to a Nazi website. Now, Autocomplete doesn’t offer that suggestion at all and typing it in directs you to the Holocaust denial Wikipedia page and the United States Holocaust Memorial Museum.
Left to its own devices, Google’s algorithm promoted Holocaust denial. It did so not because it’s a Nazi, but because it doesn’t know that such views are abhorrent. And so Google had to adjust it. It’s a curious, values-based move from a company that ordinarily presents itself as impartial. But when such impartiality is rightly called out as racist, sexist and aiding the spread of conspiracy theories, hatred and division, Google has been forced to act.
The way Google Images objectifies female celebrities, and women generally, is another facet of this problem. “Bias, fairness, explainability, accountability, and transparency requires work from researchers, policymakers, and the public,” says Caliskan. And right now, that’s not happening. As the recent furore around sexual harassment claims and an alleged culture of retaliation shows, Google is struggling to address unrest amongst its own staff let alone grapple with the challenge of admitting to, and then fixing, a technical problem of its own creating.
And, overwhelmingly, such action is reactive rather than proactive. As a result, Google has rarely been out ahead of this problem. Issues around Autocomplete – and algorithmically-generated search suggestions on Google Images – show that Google urgently needs to clean up its data. Through Google’s inaction, its algorithmic echo chamber risks reinforcing sexist views. As algorithms are put to work on more and more areas of our lives, it will become ever-more urgent that such biases are expunged.
Google claims to be impartial – its algorithm merely observes our world and mirrors it back at us. But without oversight of how its algorithm works – from the data it’s trained on to how it interprets that data to the results it spits out – Google’s definition of impartiality is hard to grasp. Drill down and the company’s search algorithm can be seen perpetuating racist and sexist views and amplifying societal biases. A Google Images search for a female celebrity suggests that they are an object to be leered over while our interest in male celebrities is apparently focussed on their profession, not their bodies. It’s yet another example of how, without scrutiny and transparency, algorithms are warping how we view our world.
From to anti-vaxxers to white supremacists, Google’s search algorithm has consistently given undue prominence to dangerous, fringe views. But the problem is wider and far more nuanced than that. Ultimately, we’re still nowhere close to understanding what world Google is showing us – and who it is turning us into.
This article was originally published by WIRED UK
