Community Notes and attempted assassinations
Community Notes and attempted assassinations
A look at how X's crowdsourced fact check feature handled the flurry of rumors spawned by a gunman's attack on Donald Trump
In the aftermath of the attempted shooting of former U.S. president and current Republican presidential candidate Donald Trump at a July 13th, 2024 campaign event, a variety of false claims regarding the attack circulated on X (the social media platform previously known as Twitter). Users of Community Notes, X’s crowdsourced fact checking feature, quickly began attempting to address misleading claims about this incident, which ran the gamut from ideologically-motivated misidentifications of the shooter to assertions that the attack was a false flag, with a generous dose of altered and misrepresented images thrown in for good measure.
On the bright side, Community Notes was successful in labeling some of the most viral false posts regarding the shooting, many of which were subsequently deleted by the post authors. However, the system continues to struggle with scale, as the rumors and falsehoods addressed in the Community Notes on popular posts continue to spread unchecked in posts from smaller accounts, particularly in cases where the posts do not contain images.
All proposed Community Notes, along with ratings and current note statuses, are freely available for download, enabling study of the usage and performance of the Community Notes system. This analysis covers notes written between 6 PM Eastern Daylight Time on July 13th, 2024 and 8 PM July 15th, and includes only those notes that contain the text “Trump” as well as at least one of “shoot”, “assassin”, “gunman”, “sniper”, “bullet”, “rifle”, “rooftop”, “AR-15”, and various conjugations thereof, for a total of 1109 notes on 881 distinct posts.
Of these, 96 (8.7%) are currently rated helpful and are displayed alongside the relevant posts, while 41 (3.7%) are rated unhelpful, and the remaining 972 (87.6%) still need more ratings from users with a variety of perspectives. The majority of the posts with visible or proposed Community Notes are from paid accounts; 445 of the 597 accounts (74.5%) whose potentially noted posts are still online have blue verification checkmarks.
Caveats: there are some shortcomings to the dataset used here, particularly that it is likely to exclude non-English Community Notes and, by extension, most notes on non-English posts. Also, the numbers in this analysis only include the initial post a Community Note was applied to, so cases where a note is shown on additional posts because they contain the same image as a post with a note are not included in the totals.
The 96 Community Notes visible at the time of analysis are attached to 83 distinct posts (some posts received more than one note). Although many of these notes were submitted the evening of the shooting, most were not shown until the following day, with a mean of ~15 hours and a median of ~12 hours between the time a note was submitted and the time it was displayed. Due to the relative lack of visible notes in the immediate aftermath of the shooting, Community Notes is unlikely to have done much to stem the initial wave of misinformation inspired by the attack. On the other hand, Community Notes was reasonably effective at getting the creators of false viral posts to remove them; 35 of the 83 posts with visible notes (42.2%) were subsequently deleted by their authors.
One of the primary shortcomings of Community Notes is that it has trouble dealing with scale. The system is very good at labeling individual false posts once they become popular, but still falls short when a false claim is repeated frequently in less-viewed posts from smaller accounts. Two good examples related to the shooting are a false claim that Trump was actually shot in the chest but was saved by a bulletproof vest, and a false claim that Trump used a “blood pill” to fake his injury. In both cases, posts from large accounts with millions of views were successfully caught by Community Notes, but obscure posts repeating the same claims continued to proliferate. This phenomenon is somewhat ameliorated in the case of image posts, due to Community Notes’ use of image matching to show notes on all instances of a debunked image from a noted post.
As is often the case with high profile acts of public violence, several incorrect identifications of the Trump rally shooter circulated on social media prior to law enforcement identifying the suspect as Pennsylvania resident Thomas Matthew Crooks. In many cases, Community Notes were added to correct these misidentifications in popular posts, but in at least one instance, an attempt was made to use Community Notes to further spread an incorrect identification. Community Notes user Notable Grass Eagle unsuccessfully attempted to add notes falsely identifying the would-be-assassin to multiple posts correctly naming Crooks as the shooter.
What conclusions are we to draw from this? It’s a mixed bag. On the one hand, Community Notes was demonstrably effective at labeling misleading viral posts about the shooting, as well as at rejecting attempts to use Community Notes itself as a vector for misinformation on the topic. The system also appears to have become more effective at getting users to delete misleading content, with roughly 42% of the labeled posts regarding the shooting being deleted. This is a distinct improvement from the state of affairs nine months ago, when only 7% of noted posts regarding the situation in Israel and Gaza wound up being voluntarily deleted.
On the other hand, Community Notes continues to struggle with scale in multiple ways. Although the addition of the image matching feature has resulted in some fact checks being propagated to additional posts, the system by and large still struggles to address the repetition of false claims by smaller accounts, even if those claims have already been addressed via Community Notes on popular posts. The latency of the system is also still too high to handle events of this sort in the initial “breaking news” phase, as the first fact checks took several hours to appear on the relevant posts.