If you Google “Was the Holocaust real?” right now, seven out of the top 10 results will be Holocaust denial sites. If you Google “Was Hitler bad?,” one of the top results is an article titled, “10 Reasons Why Hitler Was One Of The Good Guys.”
In December, responding to weeks of criticism, Google said that it tweaked it algorithm to push down Holocaust denial and anti-Semitic sites. But now, just a month later, their fix clearly hasn’t worked.
In addition to hateful search results, Google has had a similar problem with its “autocompletes” — when Google anticipates the rest of a query from its first word or two. Google autocompletes have often embodied racist and sexist stereotypes. Google image search has also generated biased results, absurdly tagging some photos of black people as “gorillas.”
The result of these horrific search results can be deadly. Google search results reportedly helped shape the racism of Dylann Roof, who murdered nine people in a historically black South Carolina church in 2015. Roof said that when he Googled “black on white crime, the first website I came to was the Council of Conservative Citizens,” which is a white supremacist organization. “I have never been the same since that day,” he said. And of course, in December, a Facebook-fueled fake news story about Hillary Clinton prompted a man to shoot up a pizza parlor in Washington D.C. The fake story reportedly originated in a white supremacist’s tweet.
These terrifying acts of violence and hate are likely to continue if action isn’t taken. Without a transparent curation process, the public has a hard time judging the legitimacy of online sources. In response, a growing movement of academics, journalists and technologists is calling for more algorithmic accountability from Silicon Valley giants. As algorithms take on more importance in all walks of life, they are increasingly a concern of lawmakers. Here are some steps Silicon Valley companies and legislators should take to move toward more transparency and accountability:
1. Obscure content that’s damaging and not of public interest.
When it comes to search results about an individual person’s name, many countries have aggressively forced Google to be more careful in how it provides information. Thanks to the Court of Justice of the European Union, Europeans can now request the removal of certain search results revealing information that is “inadequate, irrelevant, no longer relevant or excessive,” unless there is a greater public interest in being able to find the information via a search on the name of the data subject.
Such removals are a middle ground between information anarchy and censorship. They neither disappear information from the internet (it can be found at the original source) nor allow it to dominate the impression of the aggrieved individual. They are a kind of obscurity that lets ordinary individuals avoid having a single incident indefinitely dominate search results on his or her name. For example, a woman in Spain whose husband was murdered 20 years ago successfully forced Google Spain to take news of the murder off search results on her name.
“Such removals are a middle ground between information anarchy and censorship.”
2. Label, monitor and explain hate-driven search results.
In 2004, anti-Semites boosted a Holocaust-denial site called “Jewwatch” into the top 10 results for the query “Jew.” Ironically, some of those horrified by the site may have helped by linking to it in order to criticize it. The more a site is linked to, the more prominence Google’s algorithm gives it in search results.
Google responded to complaints by adding a headline at the top of the page entitled “An explanation of our search results.” A web page linked to the headline explained why the offensive site appeared so high in the relevant rankings, thereby distancing Google from the results. The label, however, no longer appears. In Europe and many other countries, lawmakers should consider requiring such labeling in the case of obvious hate speech. To avoid mainstreaming extremism, labels may link to accounts of the history and purpose of groups with innocuous names like “Council of Conservative Citizens.”
In the U.S., this type of regulation may be considered a form of “compelled speech,” barred by the First Amendment. Nevertheless, better labeling practices for food and drugs have escaped First Amendment scrutiny in the U.S., and why should information itself be different? As law professor Mark Patterson has demonstrated, many of our most important sites of commerce are markets for information: search engines are not offering products and services themselves but information about products and services, which may well be decisive in determining which firms and groups fail and which succeed. If they go unregulated, easily manipulated by whoever can afford the best search engine optimization, people may be left at the mercy of unreliable and biased sources.
“Better labeling practices for food and drugs have escaped First Amendment scrutiny in the U.S. Why should information itself be different?”
3. Audit logs of the data fed into algorithmic systems.
We also need to get to the bottom of how some racist or anti-Semitic groups and individuals are manipulating search. We should require immutable audit logs of the data fed into algorithmic systems. Machine-learning, predictive analytics or algorithms may be too complex for a person to understand, but the data records are not.
A relatively simple set of reforms could vastly increase the ability of entities outside Google and Facebook to determine whether and how the firms’ results and news feeds are being manipulated. There is rarely adequate profit motive for firms themselves to do this — but motivated non-governmental organizations can help them be better guardians of the public sphere.
4. Possibly ban certain content.
In cases where computational reasoning behind search results really is too complex to be understood in conventional narratives or equations intelligible to humans, there is another regulatory approach available: to limit the types of information that can be provided.
Though such an approach would raise constitutional objections in the U.S., nations like France and Germany have outright banned certain Nazi sites and memorabilia. Policymakers should also closely study laws regarding “incitement to genocide” to develop guidelines for censoring hate speech with a clear and present danger of causing systematic slaughter or violence against vulnerable groups.
“It's a small price to pay for a public sphere less warped by hatred.”
5. Permit limited outside annotations to defamatory posts and hire more humans to judge complaints.
In the U.S. and elsewhere, limited annotations ― “rights of reply” ― could be permitted in certain instances of defamation of individuals or groups. Google continues to maintain that it doesn’t want human judgment blurring the autonomy of its algorithms. But even spelling suggestions depend on human judgment, and in fact, Google developed that feature not only by means of algorithms but also through a painstaking, iterative interplay between computer science experts and human beta testers who report on their satisfaction with various results configurations.
It’s true that the policy for alternative spellings can be applied generally and automatically once the testing is over, while racist and anti-Semitic sites might require fresh and independent judgment after each complaint. But that is a small price to pay for a public sphere less warped by hatred.
We should commit to educating users about the nature of search and other automated content curation and creation. Search engine users need media literacy to understand just how unreliable Google can be. But we also need vigilant regulators to protect the vulnerable and police the worst abuses. Truly accountable algorithms will only result from a team effort by theorists and practitioners, lawyers, social scientists, journalists and others. This is an urgent, global cause with committed and mobilized experts ready to help. Let’s hope that both digital behemoths and their regulators are listening.
EDITOR’S NOTE: The WorldPost reached out to Google for comment and received the following from a Google spokesperson.
Google was built on providing people with high-quality and authoritative results for their search queries. We strive to give users a breadth of content from a variety of sources, and we’re committed to the principle of a free and open web. Understanding which pages on the web best answer a query is a challenging problem, and we don’t always get it right.
When non-authoritative information ranks too high in our search results, we develop scalable, automated approaches to fix the problems, rather than manually removing these one-by-one. We are working on improvements to our algorithm that will help surface more high quality, credible content on the web, and we’ll continue to improve our algorithms over time in order to tackle these challenges.
We’ve received a lot of questions about Autocomplete, and we want to help people understand how it works: Autocomplete predictions are algorithmically generated based on users’ search activity and interests. Users search for a wide range of material on the web ― 15 percent of searches we see every day are new. Because of this, terms that appear in Autocomplete may be unexpected or unpleasant. We do our best to prevent offensive terms, like porn and hate speech, from appearing, but we don’t always get it right. Autocomplete isn’t an exact science, and we’re always working to improve our algorithms.
Our image search results are a reflection of content from across the web, including the frequency with which types of images appear and the way they’re described online. This means that sometimes unpleasant portrayals of subject matter online can affect what image search results appear for a given query. These results don’t reflect Google’s own opinions or beliefs.