Elon Musk says X users fight falsehoods. The falsehoods are winning.

When Elon Musk acquired Twitter in 2022, he laid off swaths of workers tasked with moderating the platform and embraced an experimental approach: asking users to fact-check one another.

Musk has touted the crowdsourcing program, called Community Notes, as “the best source of truth on the internet.” But the majority of accurate fact checks proposed by users on political posts are never shown to the public, according to research from the nonprofit Center for Countering Digital Hate (CCDH) and a separate data analysis by The Washington Post - suggesting that the feature is failing to provide a meaningful check on misinformation.

The consequences are potentially profound. False posts on the service were recently blamed by federal officials for hindering hurricane relief. And X is poised to play a prominent role in the U.S. presidential election, a race in which Musk is a major backer of Republican nominee Donald Trump and spreading unfounded claims of voter fraud - most of which go unchallenged by his fact-checking program.

“We’ve always thought that one area where crowdsourcing was unlikely to work is in the increasingly tribal interpretation of news,” said Gordon Crovitz, former publisher of the Wall Street Journal and co-CEO of NewsGuard, a company that tracks misinformation.

“We’re living in a highly divisive era, and people on both sides are increasingly loyal to their view of facts,” he said - undermining any system, like Community Notes, that depends on users from different perspectives agreeing on political claims.

X users can volunteer to be a Community Notes contributor and, once accepted, can propose notes that debunk or add context to posts on the platform. Participants in the project vote on which notes should be attached to a post and displayed publicly. That process uses a voting algorithm that elevates only notes that receive consensus from users with a history of voting differently.

The CCDH’s analysis, published Wednesday, tracked how Community Notes responded to 283 posts that contained election claims identified as false or misleading by independent fact-checking organizations. The researchers studied only posts that had at least one note proposed by Community Notes contributors. More than 160,000 users have proposed notes in 2024 - a sharp increase from last year.

On 229 of the posts, proposed Community Notes offered accurate, relevant context, the CCDH found. But votes from Community Notes users succeeded in publicly attaching notes to only 20 of those posts. For the other 209, or 91 percent, participants didn’t reach a consensus under the Community Notes voting system - and the program didn’t provide any public context to the misleading claim.

That findings suggest Community Notes does a poor job of responding to falsehoods relating to politics, even when contributors correctly identify posts lacking context. Separate data analysis by The Post found that even when a Community Note is publicly added to an election-related post, the process typically takes more than 11 hours - by which time the content may have reached millions of users.

Only 7.4 percent of notes proposed in 2024 that related to the election were ever shown - and that proportion has dropped even further in October, to just 5.7 percent.

“Community Notes maintains a high bar to make notes effective and maintain trust across perspectives, and thousands of election and politics related notes have cleared that bar in 2024,” said Keith Coleman, vice president of product at X, who oversees Community Notes. “In the last month alone, hundreds of such notes have been shown on thousands of posts and have been seen tens of millions of times. It is because of their quality that notes are so effective.”

Coleman said academic research has shown that Community Notes are trusted more than conventional misinformation labels by people across the political spectrum. Researchers have also found that after a Community Note has been attached to a post, its author is roughly 80 percent more likely to delete it and other users roughly 60 percent less likely to share it, he said.

On Tuesday, X announced an update to Community Notes called Lightning Notes that the company said can push notes go live on a post in less than 15 minutes.

Despite such findings, the shortcomings of Community Notes rankle volunteers such as Marco Piani, who spend time crafting notes backed by reputable sources to fight misinformation on X. The 47-year-old physicist, who lives in Canada, has suggested notes on topics including coronavirus vaccines and diversity, equity and inclusion programs but said his proposed notes are rarely voted into public view.

He often watches misleading posts rack up hundreds of thousands of views while other Community Notes users spend days arguing over whether an accurate proposed note is necessary. “It is frustrating,” Piani said. “You’re trying to set the record straight on basic facts, and essentially it is lost like tears in the rain.”

Piani joined Community Notes a few years ago when it was initially launched under the name Birdwatch, before Musk acquired Twitter. Musk changed much about the platform when he took over but embraced the project, renaming it Community Notes within days. The entrepreneur positioned it as part of his mission to restore “truth” and “free speech” to a social network he said had been tilted to favor left-wing views, echoing claims from conservative lawmakers.

Posts by Republican politicians are four times more likely than those from Democrats to have a proposed Community Note approved, according to a Post analysis. That’s despite more notes being proposed on posts from Democrats overall, largely because of the huge number of notes on accounts associated with President Joe Biden and Vice President Kamala Harris.

Although researchers who study social platforms have praised Birdwatch/Community Notes as a fresh idea in content moderation, experts on online misinformation have long argued that crowdsourced fact-checking could supplement but not replace traditional moderation by professional employees. At X, Musk has reduced conventional moderation by laying people off and welcoming back accounts previously banned for repeatedly spreading harmful misinformation.

“The test of any system is, how does it work in practice?” CCDH CEO Imran Ahmed said of the nonprofit’s analysis of Community Notes. “What we found was that, in practice, it is not working.”

The power of misinformation on X has become undeniable. Last month, posts on the platform, including by Republican vice-presidential nominee JD Vance, helped propel the unsubstantiated claim that Haitian migrants were eating residents’ pets in Springfield, Ohio, to national attention. After Trump repeated the falsehood in the September presidential debate, life in Springfield was upended by bomb threats against schools.

Researchers at NewsGuard determined the claim originated in a small, private Facebook group. A Post analysis found that it spread among conservatives after being amplified on X by a verified, anonymous account called End Wokeness with 3.1 million followers. The account did not respond to a request for comment.

The false post went unchallenged by Community Notes users for four days, until one contributor proposed a note pointing out that police and city officials had debunked the claim. It cited five articles and forum posts, but the note didn’t get enough votes to be publicly attached to the post.

As of Wednesday, the status of the proposed note was still “needs more ratings,” making it visible only to registered Community Notes contributors. The End Wokeness post remained on X without any fact check and had been reshared 20,000 times and viewed some 5 million times.

One of the X users most often targeted with proposed Community Notes is the project’s loudest champion, Musk.

In July, he shared a video that manipulated Harris’s voice to make it sound like she had made disparaging comments. “This is amazing,” Musk wrote, adding a laughter emoji.

The post received at least 25 proposed Community Notes stating that the video was not authentic. But 24 other proposed notes countered that no context was needed, often arguing that the post was clearly satire. No note has been approved to be shown on the post, which as of Wednesday had 243,000 reshares and 136.6 million views.

About 1 in 10 of Musk’s posts have received proposed notes, and some critics of his stewardship of X have speculated he may have rigged Community Notes to be more lenient to his account. The Post’s analysis suggests this is false: Notes were publicly shown on 24 of Musk’s tweets this year, just under 4 percent of those that received proposed notes and a rate only slightly less than average for prominent X users.

Aditya Rao, 34, a start-up founder in Argentina, said he votes on Community Notes at least two or three times a day and recently started proposing his own. It can be frustrating to see misinformation spread while users argue about whether a note is needed, he said.

But Rao would prefer to have a diverse posse of users policing the platform instead of a centralized team within the company, to reduce bias. “I don’t think there is a better way,” he said. “Free speech … is messy. And I think this is the only way to do it.”

Some data suggests Community Notes is getting messier and potentially more frustrating for contributors. Only 79,000 of the more than 900,000 notes users wrote in 2024 have been shown publicly, and the success rate is going down, from about 10 percent last year to 8.6 percent this year, The Post found.

Alex Mahadevan, director of MediaWise, a digital media literacy program of the nonprofit Poynter Institute, said the ingenious system could work well as part of a broader moderation program - but not as X’s primary bulwark against falsehoods.

“It’s essentially ineffective,” Mahadevan said. “I mean it really just does not work.”

- - -

Methodology

The Post analyzed data from X’s public dataset of Community Notes through Oct. 27, 2024. Notes were categorized as relating to politics if they included any of the terms Biden, Kamala, Harris, Trump, vote or elect. The proportion of notes rated “helpful” by contributor votes and shown to all X users was based on notes shown as of Oct. 27, 2024, and excluded those that had been shown but later taken down because voting patterns changed.

Community Notes data published by X doesn’t include full information about the underlying posts. X post data was obtained from the National Conference on Citizenship for Republican and Democratic politicians, campaigns, organizations and party officials on a list maintained by The Post. The average rate of posts with a Community Note that was eventually shown was calculated using a large database of tweets from prominent X accounts, mostly focused on politics, that had at least 10 tweets with proposed notes.

- - -

Graphic:

https://washingtonpost.com/documents/53dfe999-6ea0-4094-8118-863e45a6e38e.pdf