We conduct the first large-scale study of image-based political misinformation on Facebook. We collect 13,723,654 posts from 14,532 pages and 11,454 public groups from August through October 2020, posts that together account for nearly all engagement of U.S. public political content on Facebook. We use perceptual hashing to identify duplicate images and computer vision to identify political figures. Twenty-three percent of sampled political images (N = 1,000) contained misinformation, as did 20% of sampled images (N = 1,000) containing political figures. We find enormous partisan asymmetry in misinformation posts, with right-leaning images 5–8 times more likely to be misleading, but little evidence that misleading images generate higher engagement. Previous scholarship, which mostly cataloged links to noncredible domains, has ignored image posts which account for a higher volume of misinformation. This research shows that new computer-assisted methods can scale to millions of images, and help address perennial and long-unanswered calls for more systematic study of visual political communication.

中文翻译：

---

Facebook 上的视觉错误信息

我们对 Facebook 上基于图像的政治错误信息进行了首次大规模研究。从 2020 年 8 月到 2020 年 10 月，我们从 14,532 个页面和 11,454 个公共群组中收集了 13,723,654 条帖子，这些帖子加起来几乎占了 Facebook 上美国公共政治内容的所有参与度。我们使用感知哈希来识别重复图像和计算机视觉来识别政治人物。23% 的抽样政治图片 (N = 1,000) 包含错误信息，20% 的抽样图片 (N = 1,000) 包含政治人物。我们发现错误信息帖子中存在巨大的党派不对称性，右倾图片误导性的可能性高出 5-8 倍，但几乎没有证据表明误导性图片会产生更高的参与度。以前的奖学金，主要是对不可信域的链接进行编目，忽略了导致大量错误信息的图像帖子。这项研究表明，新的计算机辅助方法可以扩展到数百万张图像，并有助于解决长期悬而未决的呼吁，即对视觉政治传播进行更系统的研究。