Community fact-checking can sharply slow misinformation once a warning appears, but speed determines how much of the damage has already been done. A large real-world study of X, formerly Twitter, found that displaying Community Notes was associated with a 61.2% reduction in subsequent sharing of misleading posts. Yet because many notes appeared only after posts had already spread, the reduction in total reposts across the full diffusion cycle was much smaller at 14.9%.
The peer-reviewed study, published in Nature Communications, analysed 237,180 fact-checked repost cascades created over more than 20 months, from the rollout of Community Notes on 6 October 2022 through 11 June 2024. Together, those cascades had accumulated more than 431 million reposts. Rather than asking users in a survey whether a warning might change what they would share, the researchers examined how actual sharing behaviour changed around the moment a Community Note became visible.
A rare test of fact-checking in the real world
Misinformation interventions are difficult to evaluate outside controlled experiments. Laboratory studies can test whether labels improve people’s ability to distinguish true from false content or reduce stated intentions to share it, but an intention measured in an experiment is not necessarily the same as behaviour inside a fast-moving social network.
The researchers therefore used time-series data on repost counts for Community Note-eligible posts over the first 36 hours after publication. They applied difference-in-differences methods and negative binomial regression models to estimate what changed after a helpful note was displayed, comparing the evolution of treated posts with suitable controls.
This quasi-experimental design cannot provide the same random assignment as a controlled trial, but it moves the evidence closer to a causal estimate than a simple comparison between posts that did and did not receive fact-checks. The team also conducted multiple robustness and subgroup analyses to test whether the main effect depended on particular modelling choices or types of users and content.
Sharing fell sharply after notes became visible
The central result was substantial. Once users were exposed to a displayed Community Note, subsequent sharing of the misleading post fell by an estimated 61.2% on average. The effect was statistically significant across the study’s main analyses and appeared across users on both sides of the political spectrum, including people who frequently consumed misinformation.
The behavioural response was not limited to readers deciding against another repost. The odds that authors deleted misleading posts were 94.3% higher when Community Notes were displayed than when they were not. This suggests that visible crowd-generated corrections can influence both the audience encountering questionable material and, in some cases, the people who originally posted it.
The researchers note an important feature of the platform’s design when interpreting the sharing effect. X’s recommendation algorithm does not automatically impose a visibility penalty simply because a post carries a Community Note. That makes a behavioural explanation more plausible: users who saw contextual information appear to become less willing to pass the post on, rather than the platform merely suppressing distribution through an automatic downranking rule.
The biggest weakness was timing
A 61.2% reduction after annotation sounds enormous, but it does not mean Community Notes eliminated 61.2% of all misinformation engagement. Social-media posts often spread fastest soon after publication, and fact-checking systems need time for users to write, rate and approve notes.
In the study, the half-life of reposts over the first 36 hours was 5.75 hours. Only 13.5% of helpful notes were displayed before that point. By the time most notes appeared, therefore, a large share of the reposting that could have been prevented had already happened.
When the researchers estimated the effect on cumulative reposts rather than only the sharing that occurred after a note was displayed, the system-wide reduction was 14.9%. The contrast between 61.2% and 14.9% is arguably the study’s most practically important result. It shows that a fact-check can be highly persuasive when seen while still having a much smaller overall impact if it arrives late.
Influential accounts were harder to correct
The intervention did not work equally well everywhere. Community Notes were less effective for posts from influential accounts, including verified users and accounts with large follower bases. They were also less effective for political content.
One mechanism identified by the researchers helps explain this pattern. Much of the reduction in engagement came from users who had no prior interaction with the author of the misleading post. Dedicated audiences may have stronger prior trust in a creator, politician or public figure, making an external correction less influential than it is for people encountering the author without an established relationship.
That matters for platform governance because the most consequential misinformation can originate from precisely the accounts with the greatest reach. A correction system that performs well on ordinary posts but has a weaker effect inside loyal audiences may struggle with some of the content capable of producing the largest cascades.
Why crowd-based fact-checking is attractive
Traditional professional fact-checking has important strengths, including specialist expertise and established editorial standards, but it faces a fundamental scaling problem. Investigating claims can take hours or days, while social platforms generate content continuously at enormous volume.
Community systems try to distribute that workload across large groups of users. On X, enrolled contributors can propose notes on potentially misleading posts, and other contributors rate whether those notes are helpful. Notes that meet the platform’s consensus requirements can then appear directly beneath the original content.
The new evidence suggests that this model can change behaviour at scale. It also indicates that effectiveness depends on more than whether a note is accurate or persuasive. The operational process that determines how quickly a useful note reaches the public is part of the intervention itself.
The study has important limits
The research examined one platform and one specific community fact-checking implementation. The findings therefore should not automatically be assumed to apply to every social network or every version of a crowd-based correction system. Platform interfaces, recommendation systems, user populations and rules for approving notes can all influence outcomes.
The dataset also primarily reflected the American context and the global North, with an emphasis on English-language posts. Cultural and political differences could change how users interpret community-generated corrections elsewhere. The authors identify cross-cultural testing as an important direction for future work.
Because the study is quasi-experimental rather than a randomised platform trial, residual differences between treated and comparison posts cannot be ruled out completely. The researchers used extensive modelling and robustness checks to strengthen the design, but observational platform data inevitably impose constraints on causal interpretation.
The policy lesson is about speed as much as accuracy
The findings shift an important part of the fact-checking debate. A platform can have a correction mechanism that works strongly after exposure and still allow substantial misinformation to circulate because the intervention arrives after the viral phase.
For platform designers, the challenge is therefore twofold: preserve safeguards that help communities identify useful and broadly acceptable notes, while reducing the time required for those notes to become visible. Moving too slowly sacrifices reach. Moving too quickly without adequate quality controls could create a different problem by displaying weak or inaccurate corrections.
The study does not resolve that trade-off, but it quantifies why it matters. Community Notes reduced subsequent sharing by more than half once displayed, yet their overall impact was roughly one quarter as large when the full life of the posts was considered. For misinformation systems operating at social-media speed, a correct answer delivered late can still leave most of the original diffusion untouched.
Source Information
Study: Chuai, Y., Pilarski, M., Renault, T. et al. “Community-based fact-checking reduces the spread of misleading posts on X (formerly Twitter).” Nature Communications, 17, 4070 (2026).
Published: 5 May 2026
DOI: 10.1038/s41467-026-72597-0
