Why indefinite online censorship won’t save social media from hate speech
Summary: Is online censorship inevitable on social media? Going forward, social media companies will probably need to moderate some content online to prevent dangerous speech. However, trying to moderate other kinds of online speech, such as hate speech, will likely remain an endless game of whack-a-mole.
Indeed, when social media companies have tried to moderate hate speech, it often doesn’t work out too well. What’s more, it can create further problems, especially when content moderation could look indistinguishable from indefinite forms of online censorship.
Hence, to disincentivize hate speech or other forms of objectionable bigotry online, we’ll likely need a different solution. For example, a more practical approach would include personal privacy and data protection regulations, as well as ethical design practices. This approach could help disincentive the online outrage and disinformation that can give rise to hate speech on social media.
Can online censorship stop hate, outrage, and disinformation on social media?
As discussed in the previous part of this article, it’s inevitable that social media reform will need to include some content-moderation practices.
In particular, social media companies will probably need to work closely with government authorities to prevent dangerous speech online. Dangerous speech refers to speech that incites imminent violence against individuals or groups. Typically, that’s one place where we can draw a line with any sort speech, including free speech, whether online or offline.
However, what about other forms of content moderation that don’t fall into the category of dangerous speech. For example, what about hate speech—bigoted statements disparaging particular social groups (usually based on ethnicity, orientation, gender, etc.)? Should social media companies moderate hate speech on their sites? After all, there’s certainly no shortage of hateful rage and mendacity on social media. So, why not just screen and remove it all?
Problems with online censorship
Unfortunately, when social media companies have tried to moderate hate speech or other objectionable forms of outrageous or erroneous speech online, it often doesn’t work out too well in practice, even if it sounds good in theory. What’s more, it can create further problems, especially when content moderation could look indistinguishable from indefinite forms of online censorship.
To illustrate, let’s look at some of the problems that arise when social media companies try to use content-moderation practices to indefinitely censor certain content online—in particular, hate speech or other forms of objectionable bigotry. These problems include:
- The problem of scale
- The problem of algorithms
- The problems with crowdsourcing and automating
(Note: For now, we’re putting aside the theoretical problem that it’s difficult to agree on a comprehensive legal definition of what exactly does and doesn’t constitutes hate speech. Here, we’ll focus on practical problems having to do with scale, algorithms, crowdsourcing, and automating.)
Online censorship and the problem of scale
For instance, take Facebook. To date, the site has billions of users, many who post on the platform several times a day. Consequently, the amount of new user-generated content (comments, photos, videos, shares, etc.) can easily total to an innumerable volume of posts daily.
The sheer magnitude of posts on this social networking site makes it practically impossible for Facebook to hire a team of content moderators large enough that could realistically censor any or all conceivable forms of bigoted statements online.
What about outsourcing online censorship?
Okay, given this problem of sheer scale, maybe it’s possible to outsource the content-moderation job. Indeed, Facebook has done exactly that over the past few years, hiring content moderators all around the world. While some are full-time staff members, many others are contract labor. In either case, the work they perform is consistently awful and sometimes low-paid.
Needless to say, it’s unsettling to spend an entire work week consuming endless streams of all the worst humanity has to say online.
As a result, it’s not uncommon for content moderators to come across sadistic posts showcasing domestic violence, child abuse, terrorism, murder, suicide, or animal cruelty. And that’s on top of all the endless outrage and disinformation that feeds into hate speech online. As The Verge has reported, content moderators frequently suffer severe depression and anxiety.
As expected, the chronic stress from their jobs takes a serious toll on their mental health. Many end up developing post-traumatic stress disorder, relationship problems, and trust issues.
Surely, there’s got to be a better solution than inflicting this sort of drudgery on workers. Yet, even if we ignore how this work badly damages people, we’re still faced with the problem of scale. At the end of the day, it’s practically impossible for a relatively small number of content moderators (whether full-time staff or contract workers) to screen and remove an extraordinarily large number of posts.
Online censorship and the problem of algorithms
In addition to the problem of scale, there’s another factor working against content moderators: the social media algorithms that distribute content online. The algorithms that make Facebook such an effective advertising and sharing medium also make it extremely difficult to moderate.
Think about it: The goal of Facebook’s algorithms is to ‘engage’ users—that is, monitor and manipulate what you see and do on social media, in order to keep you on social media. This goal ensures you see ads or content specifically targeted at you (a.k.a. ad and content personalization).
In effect, social media algorithms work around the clock to collect data about your personal activities online—your posts, photos, likes, shares, comments, amount of time spent scrolling and glancing at content, etc. Then, they use that information to target you with ads or content that constantly hijack your attention.
Of course, what’s exceedingly effective at hijacking attention is great deal of outrageous and erroneous ads and content. In other words, what’s good at hijacking attention is often what’s outrageous; and what’s outrageous isn’t necessarily truthful. This disproportionate amount of outrage and disinformation enabled on social media can also enables plenty of hate speech. So, content moderators may try to find and delete it all.
But there’s a big problem with that approach. The algorithms that spread content online work faster than the content moderators trying to remove it.
Doomsday machine: when algorithms outperform moderators
As Adrienne LaFrance, executive editor of The Atlantic, explained in her aptly titled article, “Facebook is a Doomsday Machine”:
Every time you click a reaction button on Facebook, an algorithm records it, and sharpens its portrait of who you are. The hyper-targeting of users, made possible by reams of their personal data, creates the perfect environment for manipulation—by advertisers, by political campaigns, by emissaries of disinformation, and of course by Facebook itself, which ultimately controls what you see and what you don’t see on the site. Facebook has enlisted a corps of approximately 15,000 moderators, people paid to watch unspeakable things—murder, gang rape, and other depictions of graphic violence that wind up on the platform. Even as Facebook has insisted that it is a value-neutral vessel for the material its users choose to publish, moderation is a lever the company has tried to pull again and again.
So, although content moderation may sound good in theory, it often fails in practice, because algorithms easily outperform content moderators. In this case, trying to censor hate speech or other forms of objectionable bigotry online becomes practically futile. As LaFrance goes on to explain:
But there aren’t enough moderators speaking enough languages, working enough hours, to stop the biblical flood of shit that Facebook unleashes on the world, because 10 times out of 10, the algorithm is faster and more powerful than a person. At megascale, this algorithmically warped personalized informational environment is extraordinarily difficult to moderate in a meaningful way, and extraordinarily dangerous as a result.
Problems with crowdsourcing and automating online censorship
Given the problems of scale and algorithms, other ways to censor hate speech on social media have been tried, including:
- Crowdsourcing online censorship to social media users
- Automating online censorship using artificial intelligence (AI)
Are either of of these realistic ways to screen and remove hate speech on social media sites? Let’s look at each option more closely.
Crowdsourcing online censorship to social media users
One option is to crowdsource part of the content-moderation job to social media users. The idea is that users can ‘flag’ content they find highly offensive or objectionable. Then, professional content moderators can step in and censor that content ad hoc. Again, it’s an appealing idea in theory. But in practice, crowdsourcing online censorship is highly unreliable.
After all, users may not comprehend the technical criteria as to what content should (or shouldn’t) get screened and removed. And asking billions of users to reach a clear consensus on such technical criteria is practically impossible. Even assuming they all can somehow establish and understand these criteria, they may not always see what content should be taken down.
Additionally, this sort of crowdsourcing is prone to abuse by social media users. For example, it could lead to an endless game of tit for tat, in which some users may wish to suppress other voices in ways that have nothing to do with their actual content. Surely, there’s got to be a more reliable solution. Which brings us to yet another idea: automating online censorship using artificial intelligence (AI).
Automating online censorship using AI
Granted, it’s possible to use AI in the form of artificial neural networks to screen and remove content online. These artificial neural networks have the power to moderate content at very large scales: namely, by deciphering billions of online speech patterns in an ongoing attempt to censor hate speech on social media. But again, there’s a problem with automating online censorship with AI. Artificial neural networks are easily prone to abuse, as well as being fooled by other AI.
As P.W. Singer and Emerson T. Brooking explain in the book Like War: The Weaponization of Social Media:
The greatest danger of neural networks … lies in their sheer versatility. Smart though the technology may be, it cares not how it’s used. … Governments of many less-than-free nations salivate at the power of neural networks that can learn millions of faces, flag “questionable” speech, and infer hidden patterns in the accumulated online activity of their citizens (Singer and Brooking, 2018, p 253).
For instance, in China, Facebook has been developing a ‘smart’ censorship program: artificial neural networks designed to censor content critical of the Communist Party. Even in freer countries such as the U.S., artificial neural networks are frequently taken advantage of by bad actors, including other forms of AI.
Consider the rise of ‘deepfakes’: artificial neural networks designed to generate deceptive content. (For example, hyper-realistic images or videos depicting someone saying or doing something that never happened.)
Online trolls at home, as well as adversaries abroad, can use deepfakes to fool not just people but also AI. For instance, deepfakes can generate “Not Safe For Work” (NSFW) pictures that include a photo of a baby, thereby fooling AI into classifying the content as breastfeeding, when really, it’s pornographic. Likewise, deepfakes could generate hate speech online, but also include other images to fool AI into classifying it as something else.
Results of online censorship: a less than stellar record
Given all these practical problems, it’s little surprise that social media companies have had a less than stellar record when it comes to implementing content-moderation practices to censor hate speech. For instance, when public intellectual Eric Weinstein tweeted about Beethoven, his tweet was screened and removed. Apparently, the reason was because Beethoven spoke German … and so did the Nazis … therefore, tweeting about Beethoven is equivalent to tweeting about Nazism … Huh???
This is me tweeting. I am tweeting about Beethoven. Who spoke German. Just like the Nazis? Something? Help me out @TwitterSafety.
In other totally unrelated news, tweet engagement has been down since I started talking about Trust and Safety excesses during the 2020 election. pic.twitter.com/zlBALnLCA0
— Eric Weinstein (@EricRWeinstein) December 17, 2020
Nowadays, there are nearly countless examples of content moderation failing, but here’s the point. Trying to use content-moderation practices to censor all conceivable forms of hate speech online doesn’t appear to work very well.
Moreover, it doesn’t really get to the root of the problem, which is the fact that many social media sites are designed to monitor and manipulate what you see and do online, with the goal of keeping you on these sites for as long as possible—namely, by hijacking your attention, harvesting your private data, and selling access to your data—in order to target you with click-bait ads and viral content.
As noted, an unintended consequence is that this design tends to incentivize a lot of online outrage and disinformation. And online outrage and disinformation happen to be incredibly effective, not only at getting people hooked on social media, but also at fueling more hate speech.
Hence, if we want to reform social media, using content-moderation practices to indefinitely censor hate speech—or other forms of outrageous, erroneous content or bigotry online—will likely remain an endless game of whack-a-mole. In all probability, it’s not going to be a practical solution, at least not in the long term.
From censorship to regulation and redesign
So, what would be a practical solution? In the next two parts of this article, we’ll discuss a more practical approach to problems like hate, outrage, or disinformation online. In brief, it involves regulating social media companies and redesigning social networking sites. This approach will need to include personal privacy and data protection regulations, as well as ethical design practices.
Together, regulation and redesign could help mitigate social media designs that can incentivize outrage, disinformation, and hate speech online.
References
Singer, P.W. and Brooking, Emerson T. (2018). Like War: The Weaponization of Social Media. Boston: Eamon Dolan/Houghton Mifflin Harcourt.