A study of 13M+ Facebook posts in the run-up to the 2020 US elections found that 23% of the sampled posts with political images contained misinformation
the most common post type, accounting for 40% of posts by US politics pages and groups. Matthew Hindman / @matthindman : Other studies have suggested that misinformation posts generate higher engagement and thus a bigger algorithmic boost. We did not find that. After controlling for page followership or group size, true and misleading posts generated similar levels of engagement. Yunkang Yang, PhD / @yangyunkang : Just a few days ago, Political Communication Forum published an article on the importance of studying visual political misinfo. https://twitter.com/.... Our study answers this call and provides the empirical evidence. Yunkang Yang, PhD / @yangyunkang : This method 1) ignores content shared by pages/ public groups which are much more visible than individual users 2) treats every link to non-credible domains as misinfo. - but propaganda outlets rarely publish only falsehoods and 3) misses the biggest content category - images. Yunkang Yang, PhD / @yangyunkang : We believe that the methods used in our article allow for investigations of identity based attacks to be conducted at unprecedented breadth and resolution. Matthew Hindman / @matthindman : Going forward, studies on the prevalence of misinformation on social media should be wary of the streetlight effect. And they should stop making sweeping claims about the (lack of) misinfo on FB or other platforms while ignoring the biggest content category - images. Matthew Hindman / @matthindman : Plotting the data found little relationship between misinformation and post engagement. https://twitter.com/... Matthew Hindman / @matthindman : We went big, collecting more than 13 million posts, from more than 25,000 of the most popular politics pages and public groups, in August through October 2020. Because FB activity is so concentrated, these pages and groups produce >95% of engagement about US politics. Matthew Hindman / @matthindman : They've measured misinformation very imperfectly — just counting links to “non-credible” outlets, and using data that conflates hyperpopular pages & big public groups with posts by individuals. Yunkang Yang, PhD / @yangyunkang : Besides misinformation, we also found troubling visual content that targeted minority groups. Many of these identity-based attacks reflect deep social inequalities. Matthew Hindman / @matthindman : We used facial recognition technology to identify political figures, and perceptual hashing (p-hash) to identify duplicate images. Expert coding of posts (both images and text) was used to classify misinformation. Yunkang Yang, PhD / @yangyunkang : We conducted the first large-scale visual misinformation on Facebook, also the first large-scale visual misinformation re: US politics on any social media. Combining expert coding with computer vision, we found that more than 20% of public image posts contained misinformation. @decustecu : Fantastic new paper. As social media become primarily visual (either static images or videos), it is imperative we focus our research on those dimensions. Text-first social platforms are quickly taking the back seat to more visual ones. https://twitter.com/... Yunkang Yang, PhD / @yangyunkang : New JoC publication on visual misinformation on Facebook w/ @MattHindman and Trevor Davis (@counteraction)! The problem of misinformation is WORSE than you think. https://twitter.com/... Jay Van Bavel / @jayvanbavel : Overall, 23% of political image posts were misleading! With right-leaning images 5-8 times more likely to be misleading. An analysis of 13,723,654 Facebook posts suggests prior estimates of misinformation exposure are significant underestimates. https://twitter.com/... Dr. Mary Anne Franks / @ma_franks : “'Democratic female public figures—especially those of color—seem particularly likely to be targeted for abuse' ... This ‘identity propaganda’ seeks to delegitimize non-white groups, exploit stereotypes, and undermine representation.” https://techpolicy.press/... @CCRInitiative Justin Hendrix / @justinhendrix : They discovered a disproportionate number of political image posts that direct hate towards groups such as women and racial minorities. “Democratic female public figures—especially those of color—seem particularly likely to be targeted for abuse.” https://techpolicy.press/... @fboversight : “While some may hold that social media companies' moderation policies have addressed the problem of misinformation over the past few years, any discussion of the issue that overlooks image posts is insufficient, say the authors.” @justinhendrix https://techpolicy.press/... Emily Bell / @emilybell : 🚨new research on visual misinformation on Facebook. Taking a vast cache of images from political pages, it shows that a. misinformation is present in 20per cent of the collected images b. Mostly from right leaning pages c. Visual misinfo is not declining https://academic.oup.com/... Justin Hendrix / @justinhendrix : A new study of images collected from Facebook pages and groups in the runup to the 2020 US election finds widespread visual misinformation that is highly asymmetric across party lines, with right-leaning images 5X to 8X times more likely to be misleading. https://techpolicy.press/... Thanks: @masonpelt
Context & Ripple Effects
This study scales up the question Jonathan Albright posed in his analysis of 250K posts and 5K political ads before the midterms — Facebook's struggle with politically motivated misinformation — but points it at images, which the researchers found make up roughly 40% of what US politics pages and groups post. It also arrives after the reassessment of Avaaz's overstated false-story estimate pushed the field toward stricter measurement, and the methodology here — facial recognition, perceptual hashing, and expert human coding across 13M+ posts from 25,000+ pages and groups — reads as a direct answer to that credibility problem.
Two findings matter most for the ongoing argument: right-leaning images were 5–8 times more likely to be misleading than others, and, contrary to what Matthew Hindman notes other studies have claimed, true and misleading posts drew similar engagement once page followership and group size were controlled for. That second result cuts against the algorithmic-boost narrative just as Meta-collaborated research was already softening claims that the algorithm changes beliefs.
First-order effects
- Facebook faces evidence that its biggest moderation blind spot sits in its most-used format — images — with right-leaning visuals flagged 5–8 times more often than others, sharpening the asymmetry its policy teams must explain.
- Hindman and Yang's engagement-parity finding directly challenges the premise, common in earlier work they cite, that misinformation earns outsized algorithmic distribution on the platform.
Second-order effects
- The documented stream of hateful imagery aimed at Democratic female public figures of color gives civil-rights advocates a concrete basis to pressure Facebook to enforce harassment rules separately from fact-checking.
- The partisan skew hands critics of platform moderation a specific statistic to dispute or weaponize, intensifying the fight over how Facebook labels content rather than merely removing it.
Third-order effects
- If engagement parity replicates alongside the Meta-collaborated findings, the regulatory case shifts away from ranking-and-amplification rules toward supply-side interventions — labeling, provenance, and removal standards.
- Auditing elections increasingly requires image-level forensics like perceptual hashing and facial recognition, extending a detection struggle algorithms have had with doctored visuals since at least the 2018 analysis of their failures; studies of this design become the template for measuring visual misinformation at scale.
The trend: Election-misinformation research is pivoting from links and headlines to visual content as the dominant vector, even as its own engagement findings erode the algorithmic-boost narrative that shaped a decade of platform-policy debate.