Facebook Study Finds Cutting Untrustworthy Posts by 70% Does Not Shift Beliefs

A Surprising Result in the Fight Against Misinformation

A new peer-reviewed study conducted with Facebook’s parent company Meta has delivered a counterintuitive finding: reducing users’ exposure to untrustworthy content by 70% did not change their political attitudes or beliefs. The experiment, one of the largest of its kind ever run on a major social media platform, challenges a core assumption behind many content moderation and media literacy efforts — that limiting the reach of dubious sources will meaningfully alter what people believe.

Over a two-month period during the 2020 U.S. presidential election cycle, researchers recruited tens of thousands of Facebook users who regularly followed politically active pages. Using a combination of algorithmic ranking changes and direct restrictions, the study sharply cut the amount of content from domains rated as untrustworthy by professional fact-checkers. Yet when participants were surveyed afterward, their views on policy issues, trust in institutions, and beliefs about election legitimacy were statistically indistinguishable from those of a control group that continued seeing normal levels of such content.

The findings were published as a working paper by researchers from Meta, the University of Texas at Austin, and other institutions. They represent a significant contribution to a long-running debate about whether social media platforms genuinely influence political attitudes, or whether they simply amplify preexisting views in an environment where other sources of information — news media, family, friends, and offline experiences — dominate.

How the Experiment Worked

The study was designed to isolate the effect of exposure to low-quality information. Researchers first assembled a list of domains that fact-checkers had repeatedly rated as posting false or misleading content, including sites that had been flagged for hoaxes, conspiracy theories, and fabricated news stories. Participants were then divided into groups: one group had their exposure to these domains reduced by about 70%, while another group had it reduced by about 40%. A third group served as a control and saw no changes.

The intervention was aggressive by real-world standards. In the 70% reduction group, users saw roughly one untrustworthy post per month on average, compared with three to four posts per month for the control group. For many participants, this meant the difference between encountering questionable content several times a week and encountering it almost never.

Despite this substantial reduction, self-reported measures of political ideology, policy preferences, and beliefs about the 2020 election showed no divergence between the groups. The researchers also found no changes in users’ overall engagement with Facebook, their consumption of news on the platform, or their levels of political participation — suggesting the intervention did not simply make people less informed or less engaged, but rather that the removed content was not moving the needle in the first place.

Do Social Media Algorithms Actually Change Minds?

The results will likely fuel skepticism about the effectiveness of algorithmic interventions as a tool for changing political beliefs. For years, journalists, policymakers, and advocacy groups have pushed platforms to demote or remove misleading content, arguing that doing so would reduce polarization, improve democratic discourse, and protect public health. This study suggests the causal chain is more complicated.

One explanation is that people’s political attitudes are formed by a wide array of inputs, and social media is only one small part of that mix. Even heavy users of Facebook also watch television news, read newspapers, talk to colleagues, and absorb information from other digital services. Cutting exposure to a handful of untrustworthy domains may simply be too narrow an intervention to overcome the broader information environment.

Another possibility is that users are not simply passive consumers. If a person is strongly committed to a political position, seeing one fewer dubious article per week may have little impact, especially if they encounter similar arguments from other sources or interpret the removal as censorship, which can reinforce their original beliefs. The researchers noted that some users in the treatment groups may have actively sought out the blocked content through other means, though they found no evidence of that on Facebook itself.

There is also the question of measurement. The study relied on survey instruments that ask participants to rate their ideology on a numeric scale and to agree or disagree with specific policy statements. These measures are widely used in political science, but they may not capture more subtle changes in the way people process information — such as increased distrust of mainstream media or greater willingness to share outlandish claims. The authors acknowledge this limitation, writing that their metrics are "designed to capture general attitudes, not moment-to-moment reactions."

Comparisons to Prior Research on Misinformation Exposure

The new findings sit uneasily alongside earlier studies that painted a more alarming picture of social media’s role in spreading misinformation. For example, research published in 2018 in Science found that false news on Twitter spread faster and farther than true news. Subsequent studies showed that exposure to Russian-linked accounts during the 2016 election reached significant numbers of Americans. These findings drove a sense of urgency among regulators and led to a wave of transparency reports and content moderation initiatives across tech companies.

But a growing body of evidence has begun to question whether mere exposure is equivalent to persuasion. A series of experiments involving Twitter, Facebook, and YouTube have found that while misinformation is indeed prevalent and often engaged with, its direct effects on individual attitudes are often small or nonexistent. A major 2023 study by researchers at MIT and Harvard, for instance, found that incentivizing Facebook users to follow ideologically diverse sources reduced their exposure to political content but did not reduce political polarization.

The current study adds to this picture by showing that reducing exposure to the most problematic content — the kind that fact-checkers have flagged — also fails to move attitudes. This does not mean misinformation is harmless. It can still waste people’s time, erode trust in reliable sources, and influence narrow beliefs about specific events, even if it has little effect on broad ideological positions. But the findings imply that misinformation’s power may lie more in animating and mobilizing people who already agree with it than in converting those who do not.

Implications for Meta, Policymakers, and the Public

For Meta, the results are both validating and awkward. On one hand, they suggest that the company’s existing fact-checking and demotion systems are not having a large negative effect on user attitudes — that is, the reductions are not making people more extreme or more disengaged. On the other hand, they undercut the argument that those systems are delivering significant societal benefits through attitude change. Meta has spent billions of dollars on integrity infrastructure, and while it can point to reductions in harmful content, this study limits the claim that those investments are changing minds.

Policymakers may also need to recalibrate expectations. Legislative proposals in the European Union, the United States, and elsewhere often assume that stricter rules on algorithmic amplification will lead to healthier democracies. The study does not show that such rules are useless — reduced exposure could still prevent rare but severe harms, such as someone acting on a dangerous conspiracy theory. However, it suggests that the average user’s political beliefs are unlikely to shift simply because a platform shows them fewer untrustworthy posts.

For the public, the study offers a humbling reminder of the limits of digital media. The belief that social media platforms are uniquely powerful agents of persuasion has become a common narrative in everything from election postmortems to parenting advice. This experiment, conducted on a massive scale with real users in a real election, found no measurable effect. That does not mean platforms have no influence — they clearly shape what people see, who they talk to, and what they perceive as normal — but it does suggest that attitudes are stickier than many suppose.

What Comes Next

The research team plans to continue analyzing the data, with future work focusing on whether the interventions affected specific subgroups, such as people with low political knowledge or those who primarily get news from Facebook. They also intend to study whether repeated, longer-term exposure to untrustworthy content has compounding effects that were not visible over the two-month window. A longer experiment, perhaps lasting a year or more, would be costly and logistically challenging, but it would answer whether the null result is a matter of duration rather than mechanism.

Other researchers are already calling for more ambitious field experiments that combine content reduction with positive interventions — for example, boosting trustworthy sources while simultaneously demoting untrustworthy ones. The current study only removed content; it did not try to enhance the quality of what replaced it. Some evidence suggests that small changes to the ranking algorithm — such as giving preference to articles rated highly by professional journalists — could have different effects, though those interventions are harder to implement without raising concerns about bias.

In the meantime, the study stands as a cautionary tale for anyone who believes that cleaning up the information environment is the key to fixing democratic dysfunction. Misinformation is real, and it is a problem. But understanding its actual effects, as this experiment shows, requires looking beyond exposure metrics and asking whether any intervention can overcome the messy, redundant, and deeply social way in which people form their beliefs.

The working paper has not yet undergone final peer review with a major journal, though it has been reviewed by academics associated with the project. Its data were provided by Meta under a transparency arrangement that allows independent researchers access to internal platform data under strict privacy protections. The findings are expected to be refined in future revisions, but the core result — that a 70% reduction in untrustworthy content had no effect on attitudes — is unlikely to change.