Meta-backed research finds exposure to ‘untrustworthy’ social media is rare. The fine print is less reassuring

How much of what you see on social media comes from sources that repeatedly post falsehoods? And if that content quietly vanished from your feed, would you believe anything different months later?

A new study published today in Science Advances offers the most rigorous answers yet. It comes from a collaboration between Meta and academic researchers who studied the 2020 United States election. It was run from inside the machine – on the live feeds of Facebook and Instagram.

The answers sound reassuring: exposure to lies was rare for most users, and removing them did not measurably shift anyone’s beliefs. But, as a researcher who studies how misinformation permeates these platforms, I think the fine print matters more than the headline.

Harder than it looks

Our political attitudes are shaped by television, talk radio, politicians, family and friends, so isolating social media’s contribution is particularly difficult. People also self-select: those drawn to fringe content surround themselves with others like them in tight clusters, so observation alone can easily miss them, and proves little.

The clean way through is a randomised experiment inside the platform, and only the platform can run one. That is what makes this study valuable.

The team of researchers, led by Olivier Bergeron-Boutin from the University of California, Berkeley, first measured how much content – across 231 million adult US Facebook accounts and 200 million Instagram accounts – came from “untrustworthy sources”. These included pages, groups, accounts and websites repeatedly caught posting falsehoods by Meta’s third-party fact-checkers.

Exposure was rare on average: 1.1% of what the median Facebook user saw and 0.1% on Instagram. But it was strikingly concentrated. For example, some 53 million Facebook accounts – 23% of users – received nearly 80% of content from untrustworthy sources.

For the most exposed 2.5%, untrustworthy sources supplied the majority (roughly 60%) of the political and social content they saw from pages and groups.

Tweaking the algorithm

For three months around the 2020 election, posts from these repeat offenders were then removed from the feeds of half of nearly 16,000 consenting users. Exposure in the targeted feeds fell by about 70%.

The researchers tracked ten outcomes on each platform, including belief in false claims, the ability to tell true from false, polarisation and trust in media. Nothing moved, not even among the heaviest consumers (though estimates for that group are less precise).

This echoes the findings of similar experiments: changing the feed algorithm, removing reshares, even paying people to log off entirely barely moved attitudes in the short run.

The grey zone

So is the misinformation debate settled? Not quite. The weak points of the study sit at the joints between what was measured and what will likely be claimed on its behalf.

A source counted as “untrustworthy” only after Meta’s fact-checking partners caught it twice. Those partners published roughly 300 US fact-checks in January 2020; matching software then spread such verdicts across far more content, about 180 million labelled pieces during 2020.

But anything the pipeline didn’t catch counted as trustworthy. “Exposure was rare” really means “exposure to detected repeat offenders was rare”.

The grey zone stays invisible by design.

In the one domain where it has been quantified, COVID vaccine content, misleading material that was never flagged had an estimated 46 times the total impact of everything the fact-checkers caught.

And in the case of state-backed influence operations that weaponise information, accounts are rotated and stay below exactly these detection thresholds.

Not normal Facebook

The control group was not experiencing everyday Facebook.

From around early November 2020, Meta deployed 63 emergency election measures aimed at reducing the flow of falsehoods for all users. One re-analysis of a companion study estimates these cut the share of untrustworthy content in ordinary feeds by around 24%. The experiment measured extra cleaning on top of an already scrubbed platform.

The dose was modest too: about three out of 250 daily posts a Facebook user viewed were removed (12 of roughly 1,050 on Instagram), for three months. This is compared to attitudes developed over decades and anchored in identity and community.

And, crucially, participants had to volunteer for a study. People who agree to participate in such experiments are not always representative of the population at large.

Twelve of the 30 authors are current or former Meta employees, and every measurement instrument is in Meta’s proprietary pipeline. The lead academics took no money from Meta and had final say over the text, but the project’s own independent evaluator previously called the arrangement “independence by permission” and warned it should not become the model.

Remembering what was actually measured

Since 2020, much has changed on social media – especially when it comes to how Meta manages misinformation and disinformation.

The company ended third-party fact-checking in the United States in April 2025. By the study authors’ own admission, the intervention they tested “is no longer possible” there, and exposure to falsehoods “could increase considerably”.

Its replacement, Community Notes, cuts resharing when notes appear but typically arrives after posts have gone viral. Even Meta’s own Oversight Board judged it inadequate as a substitute in March 2026.

There is a risk this study will be leveraged by big tech platforms to exonerate themselves – for example, by saying “users are rarely exposed to” falsehoods. It may also motivate them to stop moderation efforts on the basis that “removing misinformation changes nothing”.

But remember what was actually measured: detected sources, a scrubbed baseline, a small dose, a short window, and people willing to be watched.

The study is good science. But some conclusions drawn from it might not be.

Marian-Andrei Rizoiu receives funding from the Advanced Strategic Capabilities Accelerator (ASCA), the Australian Department of Home Affairs, and the Commonwealth of Australia as represented by the Defence Science and Technology Group of the Department of Defence. He is also the Director of the Defence Innovation Network.