(Credit: © PhotoPlus+ – stock.adobe.com)
A landmark study published in Science Advances has challenged the long-standing assumption that social media exposure to misinformation is the primary engine driving political polarization and false beliefs. Conducted during the volatile 2020 U.S. presidential election, the research followed approximately 15,900 Facebook and Instagram users to measure the impact of content moderation. While researchers successfully cut participants’ exposure to unreliable sources by roughly 70% for three months, the experiment yielded a surprising result: there were no measurable changes in participants’ political beliefs, partisan attitudes, or their propensity to believe in false claims. Even among users who were historically heavy consumers of misinformation, the reduction in exposure failed to alter their fundamental worldview or their trust in institutions.
The study provides a critical look at the distribution of misinformation, revealing that it is far less ubiquitous than the average user might assume. For the typical Facebook user, content from unreliable sources comprised only 1.1% of their feed; on Instagram, that figure was a negligible 0.1%. However, the data reveals a stark disparity in consumption. Roughly 23% of Facebook users and 11% of Instagram users were responsible for nearly 80% of all exposure to unreliable content. While the majority of the population encounters very little misinformation, a concentrated segment of users is deeply embedded in environments dominated by untrustworthy sources, which the researchers found were also significantly more likely to peddle hate speech and uncivil content.
The experimental methodology involved splitting participants into two groups: one where content from flagged, untrustworthy accounts was systematically removed, and a control group that experienced the platform as usual. Despite the technical success of the intervention—which drastically reduced the volume of “junk” news appearing in the treatment group’s feeds—the researchers observed no significant shifts in the outcomes they tracked. Whether it was distrust of mainstream media, views on COVID-19, or opinions regarding election integrity, the treatment group showed virtually no deviation from the control group. This suggests that the relationship between online information consumption and belief formation is far more complex and resistant to platform-level “sanitizing” than previous theories suggested.
The findings also shed light on the mechanics of how misinformation travels. On Facebook, the study found that large, public-facing Pages were the primary engines for disseminating unreliable information, boasting audiences significantly larger than those of private groups. In contrast, Instagram users primarily encountered such content through accounts they had manually chosen to follow, rather than via algorithmic recommendations. This distinction is vital for policymakers, as it shifts the focus of the debate: if users are actively choosing to follow these accounts, simple algorithmic downgrading may not be enough to break the cycle of exposure. Furthermore, the study notes that flagged sources were disproportionately linked to other rule violations, such as bullying and harassment, marking them as hubs for broader problematic behavior rather than mere purveyors of false facts.
These results carry significant weight as tech companies re-evaluate their moderation strategies. In early 2025, Meta moved away from third-party fact-checking in favor of community-driven notes, a pivot that makes this study particularly timely. While the researchers caution that their experiment was limited to a specific period and a specific type of enforcement, the data indicates that current moderation tactics—while technically effective at removing content—may not be the “silver bullet” for solving political radicalization that critics have long hoped for. The study suggests that policymakers and platforms must look beyond mere exposure, as simply cleaning up feeds does not necessarily “cure” the underlying beliefs of the audience.
Ultimately, the study remains a vital contribution to the field, though its authors acknowledge several limitations, including the highly polarized nature of the 2020 election and the fact that social media is only one component of a user’s total information environment. While the research offers a sobering perspective on the efficacy of source-level enforcement, it also opens the door to deeper questions about how beliefs are formed and sustained in the digital age. By proving that a massive reduction in misinformation does not immediately lead to a more consensus-driven society, the study serves as a call for a more nuanced understanding of the digital landscape, where human psychology and choice remain just as influential as the algorithms themselves.
