Beyond the Pendulum Swings of Centralized Moderation (X/Twitter, Meta, and Fact Checking)

The crazy pendulum swings of centralized moderation by dominant social media platforms is all over the news again, as nicely summarized by Will Oremus, and explored by a stellar Lawfare panel of experts. 

We have seen one swing toward what many (mostly the right) perceive as blunt over-moderation and censorship that intensified around the 2016 election and aftermath. And now, with the 2020 election and aftermath, a swing away, to what others (mostly the left) view as irresponsibly enabling uncontrolled cesspools of anger, hate, and worse. This pendulum is clearly driven in large part by the political winds (which it influences, in turn), a question of whose ox gets gored, and who has the power to influence the platforms -- "Free speech for me, but not for thee."

This will remain a disruptive pendulum -- one that can destroy the human community and its collective intelligence -- until we step back and take a smarter approach to context and diversity of our perceptions of speech. More reliance on community moderation, as X/Twitter and Meta/Facebook/Threads are now doing, points theoretically in the right direction: to democratize that control -- but is far from being effective. Even if they really try, centralized platforms are inherently incapable of  doing that well. 

Middleware as systems thinking on how to do better

Three of the speakers on the Lawfare panel were coauthors/contributors with me in a comprehensive white paper, based on a symposium on a partially decentralized approach called "middleware." That proposes an open market in independent curation and moderation services that sit in the middle between each user and their platforms. These services can do community-based moderation in a fuller range of ways, at a community level, much more like the way traditional communities have always done "moderation" (better thought of as "mediation") of how we communicate with others. This new middleware paper explains the basics, why it is a promising solution, and how to make it happen. (For a real-world example of middleware, but still in its infancy, consider Bluesky.)

As for the current platform approach to "community moderation," many have critiqued it, but I suggest a deeper way to think about this, drawing on how humans have always mediated their speech. Three Pillars of Human Discourse (and How Social Media Middleware Can Support All Three is a recent piece on extending current ideas on middleware to support this solution that has evolved over centuries of human society. The three pillars are: User Agency, Social Mediation, and Reputation. 

Toward effective community moderation

The Augmented Wisdom of Crowds: Rate the Raters and Weight the Ratings (from 2018) digs deeper into why simplistic attempts at community surveys fail, and how the same kind of advanced analysis of human inputs that made Google win the search engine wars can be applied to social media. A 2021 post and a 2024 policy brief update that.

To understand why this is important, consider what I call The Zagat Olive Garden Problem. In the early 90s, I noticed this oddity in the popular Zagat guide, a community-rating service for restaurants: The top 10 or so restaurants in NYC were all high-priced, haute cuisine or comparably refined, except one: Olive Garden. Because Olive Garden food was just as good? No, because far more people knew it from their many locations, and they were attracted to a familiar brand with simple, but tasty, food at very moderate prices, and put off by very high prices. 

Doing surveys where all votes are counted equally may sound democratic, but foolishly so. We really want ratings from those with a reputation for tastes and values we relate to (but leavened with healthy diversity on how we should broaden our horizons). That is what good feed and recommender algorithms must do. For that, we need to "rate the raters and weight the ratings," and do so in the relevant context, as that post explains.

Back to the pendulum analogy, consider how pendulums work -- especially the subtle phenomenon of entrainment (perhaps blurring details, but suggestive): 

• 
  Back in 1666, Huygens invented the pendulum clock and discovered that if two were mounted on the same wall, their pendulum swings gradually became synchronized. That is because each interacts with the shared wall to exchange energy in a way that brings them into phase.
  
• Simplistically, moderation is a pendulum that can swing from false positives to false negatives. Each conventional platform has one big pendulum controlled by one owner or corporation that swings with the political wind (or other platform influences). Platform-level community moderation entrains everyone to that one pendulum, whether it fits or not -- resulting in many false positives and false negatives, often biased to one side or the other.
  
• Alternatively, a distributed system of middleware services can serve many individuals or communities, each with their own pendulums that swing to their own tastes.
• Within communities, these pendulums are tightly linked (the shared wall) and tend to entrain.
• Across communities, there are also weaker linkages, in different dimensions, so still nudge toward some entrainment.
• In addition to these linkages in many dimensions, instead of being rigid, the "walls" of human connection are relatively elastic in how they entrain.
• The Google PageRank algorithm is based on advanced math (eigenvalues) and can treat individual search engine users and their intentions as clustering into diverse communities of interest and value -- much like a network of pendulums all linked to one another by elastic "walls" in a multidimensional array.
• Similar algorithms can be used by diverse middleware services to distill community ratings with the same nuanced sensitivity to their diverse community contexts. Not perfectly, but far better than any centralized system.

In addition, part of the problem with current community notes, and any form of explicit ratings of content, is getting enough people to put in the effort. Just as Google PageRank uses implicit signals of approval that users do anyway (linking to a page), variations for social media can also use implicit signaling in the form of likes, shares, and comments (and more to be added) to draw on a far larger population of users, structured into communities of interest and values.

Of course there are concerns that the decentralization of middleware might worsen fragmentation and polarization. While it might have some such effect in some contexts, there is also the opposing effect of reducing harmful virality feedback cascades. Consider the fluid dynamics of an old fashioned metal ice cube tray, and how water sloshing in the open tray forms much more uncontrollable waves than in the tray with the separating insert in place.

The only effective and scalable solution to social media moderation/curation/mediation is to build distributed middleware services, along with tools for orchestrating the use of a selection of them to compose our individual feeds. That too can be done well or badly, but only with a collective effort to do our best on a suitably distributed basis can we succeed.