Cutting Misinformation on Social Media by 70% Didn’t Change What People Believed, Study Shows

Researchers ran one of the largest social media experiments ever conducted during the 2020 presidential election, and what they found adds important nuance to how we think about the problem of misinformation online. Even after scientists dramatically reduced how much content from unreliable sources users saw in their feeds, the researchers found no measurable effects on the outcomes they had set out to test — at least during this specific period and under these specific conditions.

For years, the dominant fear has been that social media is flooding ordinary Americans with a steady diet of lies, conspiracy theories, and inflammatory content, slowly poisoning how people see the world. That fear is behind countless headlines, congressional hearings, and calls for tech companies to do more. But a sweeping new study published in Science Advances, involving about 15,900 adults on Facebook and Instagram split between the two platforms, suggests the relationship between seeing bad content and believing bad things is far more tangled than previously thought. Slashing exposure to unreliable sources didn’t move the needle on false beliefs about COVID-19, distrust of mainstream news outlets, partisan divisions, or views about election integrity

That doesn’t mean misinformation isn’t a real problem. The study reveals that the problem is sharply lopsided, and that simply reducing how often people encounter unreliable sources may not be the fix that policymakers and platform executives have been banking on.

Most People Barely Saw Misinformation to Begin With

One of the study’s most telling findings is just how unevenly misinformation exposure is distributed. Among a typical adult Facebook user, content from unreliable sources made up just 1.1% of everything they saw in their feed. On Instagram, that figure dropped to just 0.1% for the typical user. For the majority of people scrolling through these platforms every day, fake news and content from unreliable sources is essentially a rounding error.

But averages can be misleading. At the far end of the spectrum, a much smaller group of users was swimming in it. Among the top 2.5% of Facebook users by exposure to unreliable sources, roughly 5.8 million accounts, that content made up at least 14% of everything they saw. On Instagram, the equivalent group of about 5 million accounts saw at least 10% of their content from unreliable sources. For those heavy users, the researchers found that more than half of all the political and civic content they saw from Facebook Pages and groups came from sources flagged as untrustworthy.

Taken together, just 23% of Facebook users and 11% of Instagram users accounted for nearly 80% of all exposure to content from unreliable sources on their respective platforms. That’s roughly 53 million Facebook accounts and 22 million Instagram accounts shouldering the overwhelming majority of misinformation exposure, while hundreds of millions of other users barely encountered it.

A Massive Experiment, With a Surprising Result

To test whether this kind of content was actually changing minds, researchers ran a three-month field experiment from late September to late December 2020, right in the thick of the presidential election and its aftermath. Among the 15,928 consenting adult users enrolled across both platforms — 8,164 on Facebook and 7,764 on Instagram — roughly half were assigned to a treatment group in which content from accounts and pages flagged as untrustworthy was removed from their feeds. The other half experienced the platform as normal.

For participants in the treatment group, the intervention worked as intended on the technical level. Exposure to content from unreliable sources on Facebook feeds dropped by approximately 73%, and on Instagram’s main feed by approximately 69%. For the heaviest prior consumers of that content, the reductions were even larger. Across the study period, Facebook control participants averaged about 5.6 views per day of unreliable content, while treatment participants averaged about 2.6. On Instagram, that comparison was roughly 22.9 views per day for the control group versus about 10.6 for the treatment group.

Yet when researchers measured what participants actually believed at the end of the experiment, surveying them in November and December 2020, the treatment group looked essentially identical to the control group across the preregistered measures. Belief in specific false claims circulating on Facebook at the time, including false claims about the 2020 election and COVID-19, showed no significant change. Neither did levels of partisan hostility toward the opposing political party, trust in mainstream news organizations, or views about the fairness of the electoral system. Of the 20 preregistered outcome estimates, just one reached nominal significance — a slightly lower sense among Instagram users of being able to understand the political world — but the researchers found that result was not robust.

Researchers also checked whether the results might differ across specific subgroups: heavy misinformation consumers, strong partisans, people with low digital literacy, and others. None of those subgroups showed meaningful effects either. Even among users who had consumed the most unreliable content before the study started and experienced the steepest reductions during the study, beliefs and attitudes remained unchanged.

Where the Misinformation Was Actually Coming From

Part of what makes these findings credible is how precisely the researchers mapped the sources of unreliable content in the first place. On Facebook, the biggest driver of exposure was Facebook Pages, the public broadcast-style accounts that content creators use to reach large followings. Content from unreliable Pages made up more than half of all exposure to untrustworthy source content on the platform. Unreliable Pages had audiences that averaged more than 12 times as large as those of typical unreliable Facebook groups, giving them far more reach.

On Instagram, users primarily encountered unreliable content from accounts they had chosen to follow directly, rather than through the platform’s recommendation features. At the time of the 2020 study, Instagram’s recommendation and discovery features played only a minor role in surfacing that kind of content. This matters for debates over whether platform algorithms are primarily to blame for spreading misinformation or whether the choices users make about whom to follow are more central to the problem.

Researchers also found that the unreliable sources they identified were not just spreading misinformation in isolation. Compared with other sources on both platforms, flagged Pages, groups, and accounts were far more likely to violate platform rules, including those on hate speech and bullying. They also received more exposure to misleading content that was not outright false, and significantly more negative reports from other users.

Cutting off exposure to known bad actors on social media is technically feasible at scale, and doing so measurably reduces the amount of false, uncivil, and harmful material people see. What it does not appear to do, at least not over the course of three months during one of the most intense political periods in recent American history, is change the minds of the people who consumed that content most.

That conclusion carries particular weight right now. In January 2025, Meta announced it would stop using third-party fact-checkers and instead rely on community notes — a significant policy shift in how the platform handles misinformation. The experiment in this study depended on a separate but related mechanism: Meta’s system for identifying repeat misinformation offenders and downgrading all content from those sources. That source-level enforcement approach is no longer in place in the same form. Researchers on the project note directly that their findings say very little about what happens when fact-checking is removed entirely, only about what happened when an already-existing enforcement system was made somewhat stricter. Whether the end of these programs eventually leads to greater exposure to misinformation, and whether that exposure shifts beliefs in ways a three-month experiment could not detect, remain open and urgent questions.

---