Selected Article
For my first blog post, I selected “Algorithmic Content Moderation: Technical and Political Challenges in the Automation of Platform Governance” by Gorwa et al.
Reason for selecting the article
I chose this article because content moderation has become one of the most critical issues faced by our digitally-connected society. Platforms are increasingly used for politics, journalism, public health, activism, business reputations, individual identities, and more. Every day, millions of posts, videos, images, comments, and files are uploaded containing speech that may be illegal or harmful. To detect and remove this content, platforms use automation.
I find this topic interesting because technology isn’t neutral. Algorithms may seem “objective” or purely technical, but deciding what content to remove, block, demote, recommend, or flag is also a social and political decision. In Social Informatics, this is important because we recognize that technologies do not exist in a vacuum – they must be analyzed in context of people, organizations, values, power, and society at large.
Argument
My argument is that algorithmic content moderation is here to stay as a necessary tool for platform governance at scale, but we should be skeptical of treating it as a fully independent solution. Moderation automation can make platforms more opaque, allow unfairness to be entrenched, and obscure the politics of who can speak freely online. While automated systems can react quickly to certain types of harmful content, removing human oversight and accountability can risk free expression, and unfairly target marginalized communities.
Evidence Supporting Argument
The primary argument presented in favor of algorithmic moderation is that human moderators could not possibly review the volume of content uploaded to Facebook, YouTube, Instagram, TikTok, X, Reddit, and other platforms every day. Automated detection allows platforms to enforce policies faster, react to crises, and remove harmful content before it reaches wide audiences.
The article points to an example of Christchurch shooter’s video being uploaded to YouTube and other platforms (Gorwa et al., 2020). Automating content recognition allowed platforms to quickly identify and block many different versions of the video, something that would be impossible with humans alone screening content. In cases where harmful material is being rapidly uploaded and reposted by thousands of users, automated tools can stop its spread faster than human moderators can respond.
Additionally, automated moderation serves a clear benefit at scale for spam, child exploitation material, terrorist propaganda, cyberbullying, and copyright enforcement. Hash-matching tools may automatically screen uploads and compare them to known copyrighted files or illegal images. If there is a match based on digital fingerprints of known files, human reviewers may be bypassed to automatically flag or remove the content. In these cases, there is a clear technical benefit to automation for speed, consistency, and lack of fatigue.
Platforms also have organizational incentives to moderate content at scale. Civil society and governments increasingly expect tech companies to do something about harmful content. If companies choose not to act, they may be criticized for allowing hate speech, radicalization, child exploitation, harassment, misinformation, and more. Automated tools can serve as proof that companies are actively working on the issues.
Evidence Against Argument
While automation can improve content moderation at scale, several examples in the article highlight how automation can create new risks. First, automated moderation creates opacity. Many platforms use proprietary moderation systems that outsiders cannot review. Users are often given little information about why their content was removed, demonetized, blocked, or flagged. This can make it impossible to challenge unfair decisions.
Second, algorithmic moderation may be unfair or biased. Content moderation algorithms are created by humans and trained with data. These data sets may include linguistic and social bias against certain groups. For example, a hate speech detection system may not understand certain dialects, slang, cultural expressions, forms of political dissent, or language primarily used by minority groups. While automated systems may appear accurate on a technical level, they may harm some communities more than others. This is particularly concerning given how central social media platforms have become for free expression.
Finally, moderation automation obscures the political nature of speech decisions. Deciding whether something is hate speech, misinformation, terrorist content, harassment, or illegal is not a purely technical determination. These categories reflect social and legal norms which depend on culture, context, power dynamics, and values. Platforms would like us to think content moderation is simply an “AI” problem, but this reinforces the illusion that these decisions are made objectively by machines. In practice, humans decide what speech rules the algorithm enforces, what data it is trained on, and what outcomes are acceptable (Roberts, 2019; Gillespie, 2018).
Copyright moderation is a case study for how algorithms can create imbalanced outcomes. Automated systems may favor rightsholders, but they can also harm legitimate users, educators, commentators, reviewers, and creators that rely on fair use exceptions. Once a copyright automation tool flags a video or removes content, the uploader may have limited recourse to appeal the decision. Moderation algorithms create a power imbalance between large corporations that own copyrights and social media users.
Balanced Perspective
The most reasonable approach is to not throw out automation, but treat it as one tool in our content moderation toolbox. Automating content moderation at scale is not going away because there are too many uploads every day. However, platforms should never treat technological automation as a way to avoid accountability or transparency. Computers are fast at detecting patterns, but they do not understand context, culture, nuance, irony, or politics.
The ideal solution is a hybrid model that empowers human moderators to review difficult cases and exercise discretion. Platforms can use technology to flag potentially harmful content, but let trained reviewers make final decisions on subjective content. Platforms should also clearly explain their rules to users, provide meaningful ways to appeal decisions, publish transparency reports, and allow independent audits of their moderation systems where possible. Finally, content moderation rules should be informed by a diverse set of communities – not just engineers, executives, advertisers, and governments.
Conclusion
Content moderation algorithms demonstrate why Social Informatics is necessary. Technology both influences society and is shaped by social forces. Automated content moderation is not simply a technical challenge, it is how companies exercise power over users’ speech. Moderation technology is necessary at scale, but dangerous if left unchecked. Rather than calling for an end to automated moderation, we should demand greater transparency and accountability from companies that use it. Platforms have a responsibility to ensure automation does not come at the cost of safety, fairness, free expression, and democratic accountability.
References
Gillespie, T. (2018). Custodians of the Internet: Platforms, content moderation, and the hidden decisions that shape social media . Yale University Press.
Gorwa, R., Binns, R., & Katzenbach, C. (2020). Algorithmic content moderation: Technical and political challenges in the automation of platform governance . Big Data & Society , 7(1). 1–15.
Perel, M., & Elkin-Koren, N. (2015). Accountability in algorithmic copyright enforcement. Stanford Technology Law Review , 19 , 473. Web.
Roberts, S. T. (2019). Behind the screen: Content moderation in the shadows of social media . Yale University Press.