When Google’s Artificial Intelligence Goes Homophobic

The Tech Giant’s Alleged Biased Algorithm and Evolving Advertising Policies May Unfairly Target and is at Risk of Censoring LGBTQ Media Companies and Content Creators

Photo: Imaginima|iStock

Arjan Dijk, Google’s former VP of marketing and global executive sponsor of Gayglers (Google’s LGBTQ+ Employee Resource Group) said in a 2018 post on Google’s blog, announcing a feature allowing businesses to tag whether they were LGBTQ-friendly or provided an LGBTQ safe space, “There’s little that compares to the feeling of walking into a place and being immediately comfortable—your shoulders loosen, your breathing slows, you physically relax, knowing you can be yourself. Finding those spaces has often been hard for the LGBTQ+ community. We want to help celebrate those spaces of belonging and make them easier to find.”

The tech giant, whose mission it is “to organize the world’s information and make it universally accessible and useful” and whose tagline is “Do the right thing,” has fallen short of its professed ideals. Instead, it fiercely protects its advertising customer interests and threat to its advertising revenue, at the expense of its publishing clients, content creators and even its own LGBTQ employees. With changes to its advertising policies that have disenfranchised various LGBTQ media companies and content creators, in 2019 some of its employees demanded Google do better and stand firm behind its commitment to support the LGBTQ community.

With more than lip service, they wanted Google to show them that it, too, offered a “safe space” – not just when it’s convenient during Pride month. They also wanted Google to affirm that it wouldn’t unfairly silence, censor or discriminate against LGBTQ media companies and content creators and the safe spaces they create. This follows dismay by LGBTQ employees months prior on Google’s initial response to not penalize a right-wing YouTube commentator for his continual homophobic harassment of a gay Vox media journalist and other complaints about how Google handles sexual harassment claims. Some 20,000 of its employees internationally walked out in protest.

Admittedly, Google’s discriminatory algorithm and the black box application of its policies hit close to home. Freshfruit had been actively trying to sign up for both Google AdSense and Ad Manager (GAM) accounts for nearly two months, a process that can be approved in a matter of days or weeks. Google’s ad suite, including GAM (its ad serving technology), AdSense (a plug-and-play ad monetization solution) and Ad Exchange (a marketplace with greater publisher controls) are the favored ad products in the market for internet publishers to maintain and grow their advertising business. While previously all of these services operated as separate products on an opt-in basis, Google made the decision in 2018 to rebrand and bundle them. Therefore, in order to get a GAM account, you need to get AdSense first. While Ad Exchange is an add-on feature, it cannot be used without GAM. To partake in any of these services as a new publisher requires an application and meeting Google’s strict publisher policies.

Although hard to substantiate in the U.S., Google has roughly 90% percent market share among ad serving technology companies in the United Kingdom, which should be roughly the same prevalence in the U.S. Google also has over 80 percent market share among publisher ad networks between its AdSense and Ad Exchange products, which is how publishers make money outside of their direct advertising business. Given that online media companies primarily rely on ad spend to fund their operating costs, Google is in many ways a requisite partner to have, since over decades it has inserted itself as an intermediary between the publishing community and their direct advertisers with its offer of efficiency and yield management tools. Therefore, the vast majority of the publishing ecosystem, including LGBTQ media businesses, would find it incredibly hard to fund critical journalism without Google. Suffice to say, the inherent irony and ethical concerns shouldn’t be lost on anyone that Google both facilitates the potentiality of free speech and its censorship by acting as an arbiter via its technology and policies.

Over the past two months, Freshfruit had been back and forth with Google’s automated system and human reviewers that had flagged various policy “violations” that were not clear and were up to Google’s interpretation. As requested, we had removed any possible textual references and any images that could be deemed to be explicit, although it is our opinion that none of our content met that criteria or were unlike what you can find on other mainstream media websites. Yet, there were still issues.

Example of potentially flagged content includes “Dirty Sex” about the stigmatization of HIV after the AIDS epidemic in contrast to current considerations of the coronavirus disease pandemic, or “Blood Work” about an artist who uses human blood to protest the FDA ban against gay, bi and transgender men from blood donation. Google’s policy bans keywords that can be categorized as “explicit” and also “shocking content,” such as blood. The most suggestive images we used were of a banana covered by a medical face mask for a piece, where three subjects spoke in their words about dating during the pandemic, entitled “Love and Sex in the Time of Corona;” for the same story, an image of a man’s torso with two sets of hands hugging him; and an illustration of two gay men kissing for an advice column about the alarming number of straight women who watch gay porn—a topic also covered by Cosmopolitan magazine. Google’s response was to provide their policy for Freshfruit to review, which we combed through and found ourselves at an impasse. We didn’t see where we were not adhering to their policies and Google didn’t specify what needed to be fixed. Throughout this process, there is no human interaction. If rejected, they offer no appeals process or a way to escalate. Your only option is to post or sift through community support boards to see how others have dealt with a similar set of circumstances. However, when you are an advertiser, the opposite is true. There are ample resources, a call center, a support team, and dedicated account management if you generate considerable revenue. In fact, they invite you to contact them with your concern and they respond.

Freshfruit encountered similar issues on Facebook’s platform as well; where their algorithm blocked our website on Facebook and Instagram (as well as the same torso image), resulting in all of our Facebook content being expunged. Much of this content was already reviewed and approved to run in Facebook Ads Manager, yet its algorithm codified Freshfruit as “spam.” It took nearly one month to address through a circuitous and nebulous process, where a human being was able to observe that there was no issue and restored our content in a matter of minutes. The reason given was that it was a “false positive” flag. Unlike Google, Facebook offered a human escalation point to resolve the matter.

There have been various reports of LGBTQ content creators on YouTube who have faced similar issues, having to do with its algorithm and policies. YouTube and Google are both owned by their parent company, Alphabet Inc. Led by married couple Celso Dulay and Chris Knight, co-founders of GlitterBombTV, 12 individual LGBTQ content creators have sued the technology company for discrimination, alleging that its machine learning algorithm inhibited, outright blocked, or demonetized their publishing or advertising content or caused them financial harm in their ability to generate advertising revenue. The lawsuit alleges that Google’s algorithm and its system of human reviewers discriminated against YouTube channels that incorporated words such as “gay,’ “trans” or “bisexual” in the title. For many this meant that their content was no longer showing up in YouTube’s Explore pages or were not being recommended alongside similar videos by YouTube’s recommendation engine.

The lead plaintiffs in the case had also sought to advertise GNews, their new YouTube show – a gay version of E! Network’s Talk Soup. After attempting to build an audience by running ads about their Christmas ‘holigay’ show and lineup via Google AdWords, their ads were flagged as problematic. After inquiring with Google AdWords’ call center for the fifth time on why their 9th ad had been rejected, which Knight recorded and Freshfruit has listened to, they were told that their content was flagged as “shocking” and violated Google’s policy. The supervisor on the call explained to them that the advertisement, which he himself had reviewed and saw no issue with, did not meet Google’s standards not because of its actual content but because the show was a gay show and considered “sexual,” despite the fact that it was not about sex and had no sexual references.

“Everywhere you went we were running into walls and we were running into these obstacles,” said Dulay. “More than frustrating, it was debilitating to people’s incomes. It was keeping us from reaching our audience. People were now saying they couldn’t find us any longer. Before that, we had a lot of correspondence with people who said that we were a touchstone to the community because they lived in a place where they didn’t have any. So, we were giving them access to information and all this other stuff. So, with that sort of dwindling, we did our own website because we still wanted to keep the show going even though we were having so many issues with YouTube.”

Stephanie Frosch, an LGBTQ+ YouTuber with about 370,000 subscribers who is also affiliated with the suit told Freshfruit that she’s barely making $100 per month down from a high of $23,000 annually in 2009.

“When this first happened,” said Frosch, “I was invited to the YouTube [offices] with several other creators from very niche-specific genres from gaming to parenting to food people from the different facets of YouTube. And we were all voicing our concerns about the new algorithm and how it was affecting our monetization. YouTube’s go-to defense at the time was “It’s not just you being discriminated against, it’s other people being discriminated against.” And it’s like, wait, that makes it a lot worse. So, it’s not just queer creators. Even people of color have been voicing that when they upload content that is specific to having an open conversation about the ongoing racism in this country that they’re often demonetized. YouTube is very adamant about saying “we have queer representation and Black representation” but only if we’re not talking about the hardships they go through in our society.”

In response to their claims, YouTube’s CEO Susan Wojcicki said in a vlog interview with Alfie Deyes that, “We do not automatically demonetize LGBTQ content. We work incredibly hard to make sure that our systems are fair. We have a ML (machine learning) fairness initiative—to make sure that our algorithms and the way that our machines work are fair.”

The lawsuit currently sits with U.S. Magistrate Judge Virginia DeMarchi in Northern California. Google’s argument is that their algorithms are protected by Section 230 of the Communications Decency Act. They want the court to consider social media sites not as publishers but as “interactive computer services” that are immune to legal action in allegations of censorship. They maintain that as such a provider they can’t be sued for removing content Google considers either “obscene, lewd, lascivious, filthy, excessively violent, harassing or otherwise objectionable, whether or not such material is constitutionally protected,” so long as removal is being done in “good faith.” The U.S. Justice Department intervened to defend the statute while it simultaneously disputes the same argument in a case over censorship involving President Trump and Twitter. The president revoked Section 230 by executive order to say that Twitter can be considered a publisher. Dulay and Knight’s attorneys are using the president’s executive order to defend their position. The same attorneys are representing four Black YouTube creators in a similar lawsuit against Google.

Only upon sharing with Google that we planned to report on our experience after two prior attempts did we get a response from their press office, stating that our prior messages had gone to their junk folder. Within fewer than 24 hours our AdSense account was approved – we had made no changes to our site. When asked what the reason for the blockade had been, John Brown, their publisher policy education lead said, “Our investigation found this was an error and result of over-flagging on a few images utilized, which were incorrectly labeled. The articles’ content was never flagged by our enforcement nor found to be policy-violating.” When pressed about how the images were labeled, he said, “The automated review, as well as the initial human review labeled the images as sexual.”

Brown said he could not provide details about their review process, did not respond to my question about whether the human reviewer had not noticed the “over-flagging” but said that “Sexual/adult content/themes tend to surface more gray area cases than many other policies, as sometimes context is key. With all human review, if you have multiple people looking at the same gray area content/images, you can generate different responses.”

When asked about whether the review team included LGBTQ individuals or what efforts are made to have a diverse team of reviewers, he said “I don’t disclose any data about the members of our team externally. However, we’ve set up an external website on our diversity efforts which you can review.”

He concluded, “Our policies and enforcement aim to strike a delicate balance – we want to provide room for publishers to engage and report on topics and their opinions, while ensuring we are instilling trust in our advertiser partners. As a result, we do see some over-flagging like this. We’re working closely on recalibrating with our team to reduce such occurrences.”

There have been a number of other reported incidents where Google’s machine learning has gotten it wrong. Google demonetized The Atlantic for publishing Martin Luther King’s Letter from a Birmingham Jail on Martin Luther King Day because it violated their “dangerous or derogatory content” policy for containing the N-word. Slate Media has said that over 10 articles, covering topics such as white supremacy, slavery and hate groups have also been demonetized for the same violation because of quoting racial slurs, including the title of a book. And Buzzfeed had monetization pulled from articles about race and the LGBTQ, including a list of pioneering queer people of color.

In the midst of this, there has been a proliferation of neo-Nazi and extremist content on social media, largely on YouTube. While both Google and Facebook made efforts to remove extremist content in 2019, Google allowed several to remain untouched and uncensored for a year like those of white nationalist Richard Spencer, the former leader of the KKK David Duke and the National Policy Institute, a white supremacist group run by Spencer. Spencer’s and Duke’s accounts have now been banned. Still neo-Nazi content continues to flout YouTube’s ability to block them and they are more than likely being monetized. PewDiePie, YouTube’s most popular user with over 100 million followers, the majority of whom are impressionable teenagers, has a history of cavorting alt-right culture, such as promoting an anti-Semitic YouTube channel and making anti-Semitic jokes that many observers found disturbing.

Lest anyone say that the above shows that rather than being discriminatory against LGBTQ content, Google’s machine learning algorithm is just overly conservative and context-oblivious, I suggest they are not mutually exclusive. Let’s not forget they incorporate human reviewers. The presence of bias in Google’s machine learning models is as biased as the source of the data they evaluate, the contents of the dataset and the training of the model itself, which may include the influence of human bias.

Just this past week, Google’s star artificial intelligence ethics researcher, Dr. Timnit Gebru, was terminated from her position because of pressure from Google higher ups to retract her research paper that she was scheduled to present at a computer science conference next March. In it, she highlighted risks in Google’s large language processing models that have more representative datasets from wealthy countries or populations that have more internet access than those that do not. This results in more homogeneity in AI-generated language.

One area she discusses that this exists and why it can be highly problematic, is that large language models are heavily reliant on text and data collected from the internet. Aside from the challenges of being auditable for bias, according to a senior AI reporter at MIT who was able to access the censored research paper, Gebru makes the argument that Google’s AI presents a risk, for example, for racist, sexist and other abusive language to be normalized at the expense and nuance of newer vocabulary with less online prevalence that might be anti-racist or anti-sexist that have come out of the #MeToo and the Black Lives Matter movements. In other words, terms that are normalized may be ignored by Google’s artificial intelligence and non-normalized words may not, which may impact a host of things, including what content can be found in search or recommended, how they vet new content creators against their policies and how they block content already within their network, depending on how those language models are applied. It raises the question of whether similar biases may exist in their models with an overrepresentation of heteronormative content that could be normalized at the expense of LGBTQ content that is not. Attempts to reach Gebru to confirm this theory went unanswered.

Perhaps unscientific, Freshfruit observed ad calls from other mainstream news websites with similar content as those more likely to have been considered problematic on freshfruitmag.com—sometimes about the same subject matter. An “ad call,” is when a user’s browser makes a request for an ad from an ad network, in this case, from Google Ad Exchange or AdSense and served to the webpage by an ad server (Google Ad Manager). Naturally, we found that Google had no issue monetizing the websites Freshfruit examined. We found Google advertising in New York Magazine’s Sex Diaries section, even on an article about seducing a married dad that included multiple references to the slang term for the male genitalia. We found Google ads running on Slate’s Dear Prudence advice column that covers topics ranging from smoking marijuana, having a sexual fling and concerns about getting an erection. Google had no problem monetizing a Vice Media story similar to Freshfruit’s about women who enjoy watching gay porn, or another of Vice’s stories about helping make a porn star out of a well-endowed young man, which is riddled with sexual words.

It can be argued that these are not clear-cut examples of bias, given that the admixture of keywords and images used are dissimilar, even if thematically they are the same—It’s unclear what of Freshfruit’s editorial was considered offensive since they did not share specifics—but it raises a question about the consistency of Google’s algorithm and the data inputs by which it is informed, if even having removed everything that could be potentially considered in violation, Freshfruit was still unable to pass the review process. There is also the question of money. While Google was blocking ads from the YouTube plaintiffs in the lawsuit against them, they were also running ads on other LGBTQ content creators in the top 1% of subscribers and views committing similar violable offenses to those outlined in the plaintiff’s arguments, suggesting that perhaps a blind eye is shown if the content creator is of economic worth to Google.

It’s equally important that we protect the right to freedom of speech as it is to protect the right for having queer media and queer content. To not do so would disenfranchise a minority population due to the opinions, input or standards of the heteronormative majority. Google should decide if it wants to be in the business of blocking free speech and condoning unfair and discriminatory practices to line its own pockets or do the right thing by making the world’s information truly accessible and creating a safe space for queer businesses, queer consumers and its very own queer employees. It should release any data it has on the rate of blocking and demonetization it conducts on LGBTQ content as compared to non-LGBTQ content. It should make efforts to include a diverse group of human reviewers to mitigate anti-LGBTQ bias. It should seek to ensure that its machine learning models represent a more diverse dataset, if they truly cannot read context. Finally, Google should improve its customer application, appeals and escalation processes on the publisher side to include human-to-human communication. Then, perhaps, it can show a commitment to free speech and allow LGBTQ folks more places online to relax their shoulders, get comfortable and breathe.

“If you look at media and how it’s going,” summarized GlitterBombTV’s Knight. “Everything is streaming now, and everything is digital. If you cut a whole swath of 10 – 20 percent, however many of them we are, you’re basically silencing the voice of the LGBTQ community. They can’t get their message out. They can’t view their own content. They might be thwarted or throttled, and it doesn’t seem right to us. So, we’re trying to get LGBTQ rights into the digital age otherwise we’re going to be cut out from it.”

Leave a Reply

Read More

Sign Up

Create an account for access to events, deals and perks.

Already have an account? Log in here.

Subscribe to Our Newsletter

Get the latest scoop
fresh & fruity!

By clicking “sign up,” you agree to receive emails from Freshfruit and accept our terms of use and privacy policy.

welcome.

Log in to access events, deals and experiences.

Freshfruit

Log in to your account

Don’t have one? Sign up here.