Skip to Content

Free Expression

The Thorny Problem of Content Moderation and Bias

By Brenda Dvoskin, CDT Summer Intern and S.J.D. Candidate at Harvard Law School

In June, Senator Hawley introduced a bill to limit immunity granted by Section 230 of the Communications Decency Act to platforms that moderate content in a “politically neutral” fashion. The so-called Ending Support for Internet Censorship Act would require companies exceeding a certain amount of users or a certain threshold of annual revenue to get a certification from the Federal Trade Commission (FTC) that the company does not “moderate information provided by other information content providers in a politically biased manner.” The bill specifies that moderation is “politically biased” if it “is designed to negatively affect a political party, political candidate, or political viewpoint,” or “disproportionately restricts or promotes access to, or the availability of, information from a political party, political candidate, or political viewpoint,” or “an officer or employee of the provider makes a decision about moderating information provided by other information content providers that is motivated by an intent to negatively affect a political party, political candidate, or political viewpoint. 

The proposal was rightly criticized for a number of reasons. The bill assumes that political neutrality exists, despite the fact that there is no objective standard to define this concept. Also, politically appointed commissioners of the FTC would be the ones deciding if political moderation is in fact neutral. This proposal could create a huge disincentive to moderate, which could lead to a proliferation of misinformation and diminished usefulness of many online services. 

Hawley’s proposal emerges in the context of widespread confusion about how platforms work and, more importantly, about how we want them to work. What do we want content moderation to look like in the digital age, what are the current problems with moderation on large platforms, and how do we start to address them?

 What do we want social media networks to look like?

The power of the internet to promote freedom of expression lies in its possibility to foster different communities and empower voices that are not heard in traditional mass media or elsewhere. Some spaces may choose to be inclusive of as many viewpoints as possible, while others may function as safe spaces to encourage specific conversations. Each one of these purposes will call for moderation rules that support the creation of a specific environment. Mandated neutrality would make it harder for communities to shape their own house rules.

Even sites that intend to host a wide range of viewpoints will nevertheless typically need to employ content moderation to keep their services generally useful to their users. Non-neutral moderation is what enables them to eliminate content that, although lawful in the United States, can have a silencing effect on users who are usually the target of hateful or violent speech. Section 230 guarantees that platforms need not be neutral to misogynistic, racist, anti-Semitic or other kinds of speech that inhibit the free expression of others in order to retain legal immunity from liability for how they moderate content. 

What is the problem today?

As a handful platforms wield more power over public discourse, the public may have its own legitimate concerns and expectations about how these platforms should work. As these spaces connect billions of users, they become a hub for bad actors seeking to leverage that reach to spread disinformation, foment discord, or incite violence. But even if people around the world agree that they are dissatisfied with current content moderation practices, they emphatically don’t agree on exactly what speech should be suppressed, filtered, or downranked. There are limits to what governments can restrict, and everything else falls into this area of “difficult” speech, where users from different cultures, languages, and backgrounds have very different expectations.

The recent debate about what social media companies should do with the doctored video of Nancy Pelosi illustrates this point. The altered video showed slowed-down footage of Speaker Pelosi speaking, making her appear to be inebriated. Responses to the video varied across platforms, and it became clear that there were competing expectations of how sites should deal with this video, and with manipulated content more generally. While some were notably frustrated by Facebook’s decision to keep up the video, for example, others celebrated that the company would not take down constitutionally protected speech and would refrain from being the arbiters of truth. These same tensions and competing expectations emerge with all “difficult” speech including nudity, violence, and hateful speech.

In the meantime, companies are making content moderation calls thousands of times a day (or more). But it is often not clear to their users how they are making these decisions or what the implications are. When we are talking about gigantic platforms that have an enormous influence on the information environment, the concerns get exacerbated. As we hear platforms promising an inclusive and fair space to share content, people are legitimately interested in holding them accountable to that promise. But without access to data or independent peer-reviewed research about the impact of content moderation policies, it can be difficult to conclusively evaluate whether they have a disparate impact on particular viewpoints–another form of “data void” that can be exploited by conspiracy theorists.

A path forward

Some social media companies have begun sharing more information about their content moderation practices with the public and some scholars have started to gain better access to the inside of the decision-making processes. But there is still a long way to go. CDT has joined other human rights organizations in advocating for companies to implement the Santa Clara principles, which provide guidelines for reporting, notice, and appeal processes. Besides numbers and statistics, we would benefit from a better understanding on how individual decisions are made. A good practice would be for companies to release case studies or examples of how they have applied their policies–a kind of common law interpretation of their rules–that would help users understand what elements (what values, what technological affordances and limitations, what other practical considerations) are factored into these decisions. Making more data accessible to academics and researchers would ensure that transparency around content moderation does not come only from the companies themselves, and would help the public get trustworthy information from independent sources. 

We will probably never see all people–or even all users of a website–agree about how to address “difficult” speech on platforms, but access to better information about how content moderation works, and its influence on public debate, would enable us to have more informed conversations about what the real problems are and how we want to address them.