Cutting dubious social media sources did not shift voter beliefs

A field experiment reducing untrustworthy feeds by 70 percent shows entrenched political attitudes resist short-term online interventions.

A common assumption in modern politics holds that false and misleading internet posts actively warp the beliefs of voters. In that view, exposure acts like an infectious agent: someone encounters a fabricated claim on their social feed, absorbs it, and alters their political stance as a direct result. If that causal model were complete, turning down the volume on suspect sources would reliably reverse the damage, leading voters back toward more moderate or accurate positions. Yet human political convictions rarely change simply because an algorithmic dial moves.

Why this assumption fails matters for any society attempting to manage online political polarization. When regulatory bodies, civic groups, and technology platforms spend resources attempting to suppress low-quality material, they often assume that ordinary citizens are impressionable bystanders whose views passively reflect whatever appears on their screens. If political identities are instead deeply rooted in personal history and community ties, feed adjustments will do little to alter ideological divisions. Knowing whether digital interventions can actually shift real-world beliefs is therefore essential for evaluating whether content curation can heal democratic divides.

For an online post to alter a person's worldview, a strict sequence of events must occur. First, the person must encounter the post on their feed. Next, they must read and interpret the message rather than scroll past it. They must then find the claim credible enough to integrate into their understanding of current events. Finally, that new information must outweigh the years of social habits, family loyalties, and personal loyalties that anchored their political identity in the first place. Because the chain depends heavily on that final, resistant link, even a drastic drop in online exposure does not guarantee a change in conviction.

That disconnect was demonstrated during the 2020 United States presidential election. A study published in Science Advances, led by Olivier Bergeron-Boutin, a doctoral student in political science at the University of California, Berkeley, and co-authored by Brendan Nyhan of Dartmouth College, Jaime Settle of William & Mary, Emily Thorson of Syracuse University, and Magdalena Wojcieszak of the University of California, Davis, found that cutting exposure to untrustworthy sources on Facebook and Instagram by about 70 percent produced no measurable changes in political beliefs or attitudes.¹²³ Even among individuals who regularly saw such content prior to the intervention, reducing questionable material left their broader political outlook essentially unchanged.¹³

How did researchers test feed changes during an election?

Researchers tested the effect by quietly altering feed algorithms for consenting users over a three-month period surrounding the vote.¹³ The project grew out of a large academic collaboration with Meta, the parent company of Facebook and Instagram, designed to evaluate platform impacts on democratic processes.¹⁴ Independent researchers were granted access to internal platform data and the ability to run controlled experiments, with the academic authors retaining final say over the findings and publication decisions.⁴⁵

To separate users into distinct groups, the study relied on an established fact-checking mechanism that Meta operated at the time. When third-party fact-checkers repeatedly flagged content from particular web domains, Facebook pages, or Instagram accounts, the platform assigned those accounts negative marks.¹³ The researchers used those records to define untrustworthy sources, creating an experimental intervention that systematically demoted posts from repeat offenders.¹³

Participants in the trial agreed to take surveys and share their feed records.⁴³ The system assigned them either to a normal browsing experience or to a restricted feed where posts from repeatedly flagged sources were suppressed.¹² Across the three months of the study, that treatment reduced the volume of untrustworthy content appearing in participants' feeds by roughly 70 percent on both platforms.³¹ Yet when the researchers evaluated survey responses covering polarization, candidate evaluations, and ideological extremity, the intervention had produced no detectable change.⁴³

The study also examined the natural flow of information across active adult users in the United States and found that problematic material was heavily concentrated.³¹ Most users rarely encountered untrustworthy sources during their normal browsing. Instead, about 23 percent of Facebook users and 11 percent of Instagram users accounted for roughly 80 percent of all exposure to untrustworthy sources on the platforms.¹² The vast majority of citizens were simply not reading or seeing questionable political claims in substantial quantities.

Why did cutting untrustworthy sources leave voter views unchanged?

Voter attitudes remained stable because political identities are anchored by community ties and long-term habits that easily outlast temporary platform interventions. Most adults form their partisan loyalties early in life through family, regional culture, and enduring social networks. When an algorithm suppresses a dubious headline, it removes a single digital item, but it does nothing to dismantle the underlying social affinities that made the reader sympathetic to the claim in the first place.

Tim Tangherlini, a computational folklorist at the University of California, Berkeley, who was not involved in the research, noted that narratives do not move randomly through a society.¹² Instead, they flow across existing social groups that are already primed to receive, retell, and amplify them. While a platform can adjust what flows into a digital network, it cannot change the real-world social fabric or the shared beliefs circulating within that community.¹² He compared shifting established beliefs to steering an oil tanker, an effort where course corrections take substantial time to manifest.²¹

Furthermore, the concentration of exposure suggests that the small fraction of users who consume large quantities of dubious material are not passive victims of platform recommendations. Magdalena Wojcieszak, a communication professor at the University of California, Davis and the University of Warsaw, pointed out that individuals who seek out such content often hold well-formed, intense political views.⁵ Rather than being trapped in an accidental information bubble, they actively seek out narratives that reinforce their existing worldview, making them exceptionally resistant to a brief, feed-level intervention.⁵

What can these results not explain about modern platforms?

These findings come with meaningful boundaries and cannot serve as an assessment of how social platforms operate today. The data reflect user behavior and platform designs specific to the autumn of 2020. In 2025, Meta ended its third-party fact-checking program in the United States and retired the strike system that made this specific source-level filtering experiment possible.¹² Because platform policies and ranking systems have shifted toward crowd-sourced models like Community Notes, the study cannot describe current levels of exposure or their potential effects.¹²

The experiment also evaluates short-term interventions across a three-month window rather than cumulative exposure over decades. Participants spent years consuming media from various television, radio, and digital outlets prior to the 2020 study. Suppressing dubious posts for 90 days cannot reveal what would happen if a person spent their entire adult life insulated from low-quality political claims. Nor does the study measure the offline behavioral consequences that might occur among small, highly motivated fringe groups who consume extreme amounts of disinformation.

Finally, the authors caution that the findings should not be interpreted as evidence that misinformation carries no societal risk. As Bergeron-Boutin observed, while simple feed interventions do not shift broad attitudes, structural features of American political life continue to sustain partisan conflict.¹ The research demonstrates the operational limits of platform-level curation, but it does not evaluate whether unchecked digital falsehoods might erode institutional trust over generational timescales.

Where should public attention focus next?

Future inquiries will need to concentrate directly on the small minority of high-volume consumers rather than assuming every citizen faces identical digital risks. Brendan Nyhan emphasized that because exposure to dubious content is heavily concentrated, any resulting harms are likely concentrated as well.¹² Public attention, he argued, would be better directed toward understanding what motivates heavy consumers of extreme material and what actions they may take, rather than worrying about the average voter who spends little time thinking about politics.¹²

Gathering that evidence, however, is becoming increasingly difficult as technology companies restrict academic scrutiny. The 2020 project represented an unusual arrangement where outside researchers secured direct access to proprietary platform data under strict academic freedom agreements. As Bergeron-Boutin noted, social platforms have become significantly less transparent since that election, making similar large-scale public experiments unlikely to be replicated in the future.¹² Without structured access to commercial platforms, independent scholars will struggle to verify how future algorithmic changes shape political life.

Save this story

The record

24 sentences trace to 5 sources.

All 5Documents 1Records 0Reporting 4People 0

1. 1 S scientificinquirer Document · scientificinquirer.com · 19 Oct 2026 Cutting Facebook and Instagram Users’ Exposure to Dubious Sources by 70% Didn’t Change Their Political Views See the source
2. 2 L ls.berkeley.edu Third-party reporting · ls.berkeley.edu Does misinformation on Facebook and Instagram change what you believe? See the source
3. 3 G gsb.stanford.edu Third-party reporting · gsb.stanford.edu Untrustworthy Sources on Facebook and Instagram in 2020: Concentrated Exposure but No Attitudinal Effects See the source
4. 4 N news.utexas.edu Third-party reporting · news.utexas.edu · 27 Jul 2023 First Findings Shed Light on Role of Social Media Algorithms in 2020 Election - UT News See the source
5. 5 E en.uw.edu.pl Third-party reporting · en.uw.edu.pl Content exposure on social media | University of Warsaw See the source

Also consulted: 1 page across 1 site, none cited.