The Outrage in Your Feed Is a Governance Choice, Not a Content Problem

Based on the research ofCarmelo Cennamo and Jovana Karanovic, "Are Social Media Platforms a Threat to Democracy? An Ecosystem Governance Perspective," Journal of Management Studies, 2026

Every major platform response to the "social media is bad for democracy" problem has taken the same shape: more fact-checkers, more takedowns, more labels, more content-moderation staff. Carmelo Cennamo and Jovana Karanovic, in "Are Social Media Platforms a Threat to Democracy? An Ecosystem Governance Perspective," argue that this response treats the wrong layer of the system. They reconceive social media platforms as managed ecosystems that allocate and sell a scarce resource, user attention, among users, advertisers, and creators. In that framing, misinformation, fake news, and polarization are not accidents that content moderation cleans up after the fact. They are negative externalities of the rule system itself: algorithmic curation, network effects, and engagement-based monetization that reward whatever keeps people scrolling, arguing, and reacting. Moderation, in their account, is the platform managing a symptom it has already been paid to produce.

The Mechanism, Not the Content

The move inverts the usual debate. Most commentary about platforms and democracy asks whether a specific piece of content, a conspiracy video, an influence campaign, a doctored image, should be removed, labeled, or left alone. Cennamo and Karanovic ask a structural question instead: what does the ranking and monetization system reward, regardless of which post wins on any given day? Their answer is that engagement-optimized recommendation and advertising markets built on attention carry a built-in bias, because outrage and conflict reliably outperform accuracy at holding attention. Remove any one piece of harmful content, and the ranking system simply promotes the next post that trips the same reward function. Because the paper is conceptual and theory-building rather than empirical, it reports no new data of its own; its evidence is that the pattern it describes has already been documented, in unusual detail, inside the platforms themselves.

What Facebook's Own Engineers Found

In 2018, Meta (then Facebook) overhauled its News Feed ranking system around a new goal it called "Meaningful Social Interactions," or MSI, intended to reverse a decline in commenting and reward posts that sparked back-and-forth engagement between friends and family. As part of that overhaul, the algorithm gave emotional reactions, including the "angry" reaction, introduced in 2015 alongside love, haha, wow, and sad, a score five times that of a plain "like," according to internal documents reported by The Washington Post and summarized in Wikipedia's account of the 2021 leak. The mechanism did what it was tuned to do: it rewarded whatever provoked a reaction. Facebook's own researchers later confirmed that posts drawing angry reactions were disproportionately likely to be toxic, polarizing, or low-quality, but the company did not zero out that extra weight until September 2019, after employees had raised the issue internally for months.

The Wall Street Journal's Keach Hagey and Jeff Horwitz laid out the mechanism in their September 2021 Facebook Files report, "Facebook Tried to Make Its Platform a Healthier Place. It Got Angrier Instead." Internal memos they reviewed showed the 2018 change rewarded outrage and sensationalism rather than the warmer, family-oriented interactions it was designed to promote, and that CEO Mark Zuckerberg resisted proposed fixes when they risked reducing engagement. Frances Haugen, the former Facebook employee who supplied those documents and later testified before the U.S. Senate on October 5, 2021, put the choice in blunt terms in the official hearing record: "Facebook is aware that choices it made in establishing... 'meaningful social interactions,' so engagement-based ranking... didn't care if you bullied someone or made hate speech in the comments that was meaningful," and that "Facebook became a $1 trillion company by paying for its profits with our safety, including the safety of our children." Senator John Thune, in the same hearing, summarized the Journal's reporting for the record: Facebook "altered its algorithm in an attempt to boost these meaningful social interactions... but rather than strengthening bonds between family and friends on the platform, the algorithm instead rewarded more outrage and sensationalism."

None of this required a bad actor posting false content. It required a ranking rule that could not tell "meaningful" from "enraging," because engagement was the only signal it was built to read.

Facebook became a $1 trillion company by paying for its profits with our safety, including the safety of our children.

When a Platform Tried the Other Lever

If the mechanism runs through ranking design rather than content, the more informative test is what happens when a platform changes the ranking itself, without removing any content. YouTube ran close to that test. Starting in January 2019, it began adjusting its recommendation system to reduce how often it suggested "borderline content", material that brushed against, but did not clearly violate, its policies, such as videos promoting conspiracy theories or medical misinformation. By its own account, published on the YouTube Blog in December 2019, the company had rolled out more than 30 separate changes to its recommender system, none of which removed the videos in question. The result, YouTube reported, was a 70 percent average drop in watch time driven by non-subscribed recommendations of that content in the United States, comparing a late-2019 period against the January 2019 baseline. The videos still existed. The moderation queue was not the lever that moved. What changed was what the recommendation algorithm chose to surface.

That is precisely the distinction the paper's framework predicts should matter: the same content can persist on a platform while its downstream reach, and its capacity to generate misinformation exposure, is governed by the ranking system's design.

Regulators Are Starting to Separate the Two Problems

The European Union's Digital Services Act, fully applicable across the EU from February 17, 2024, and already binding on the largest platforms from August 25, 2023, offers the clearest evidence that this distinction is moving from academic argument into binding law. Under Article 34, providers of "very large online platforms" and search engines, those reaching more than 45 million monthly users in the EU, about 10 percent of the Union's population, must identify and assess "systemic risks" stemming from "the design or functioning of their service and its related systems, including algorithmic systems." Article 35 then lists the mitigation measures available, and it is telling how the list is structured: item (c) covers "adapting content moderation processes," while item (d), a distinct and separate measure, covers "testing and adapting their algorithmic systems, including their recommender systems." The regulation does not treat ranking design as an input to moderation; it treats it as its own governance object, with its own risk assessment and its own remedy. The European Commission's enforcement guidance notes it can "order access to the VLOPs' data and algorithms, e.g. to assess how the algorithm/recommender system of a platform promotes illegal content", a power aimed at the ranking mechanism, not the moderation queue.

This is, in effect, the regulatory analogue of the paper's thesis: the recommender system and the moderation system are separate governance problems, and auditing one does not substitute for auditing the other.

The Rule for Anyone Governing an Attention Market

Cennamo and Karanovic's contribution is not a policy prescription so much as a diagnostic: before asking whether a platform is moderating content well, ask what its ranking and monetization rules are optimizing for, because that answer predicts the pattern of harm regardless of how aggressively any individual post gets reviewed. The Facebook Files and Haugen's testimony show what happens when a company discovers, through its own research, that its ranking choice is generating the externality and treats that as a cost of doing business rather than a flaw to fix. YouTube's borderline-content work shows the same lever run the other way, with a measurable effect on reach that no takedown could have produced as efficiently. The DSA shows a regulator writing that distinction into law, treating recommender-system design as accountable in its own right.

None of this makes content moderation worthless; illegal content still needs to come down. But the paper's real claim is a ceiling on what that layer can do: a moderation queue can only catch what a ranking system has already decided to amplify. If the reward function underneath keeps pricing outrage above accuracy, moderation is bailing water on a boat still taking it on. The governance choice that matters most happens earlier, in what the recommender is trained to maximize and what advertisers are charged to reach it, and that is where any real fix will have to be aimed.