Timeliness and Reliability in Community Notes: Accelerating Crowdsourced Fact-Checking for Visual Misinformation

Community Notes is a crowdsourced fact-checking mechanism introduced by X (formerly Twitter) that allows contributors to add corrections to potentially misleading social media posts, thereby reducing their propagation and virality. The longer a note takes to get published, the less effective it is at counteracting the spread of misinformation. In this context, visual misinformation, consisting of images or videos, sometimes paired with text, that contain false or misleading information, has evolved and increased along with the increasing presence of generative AI content. Given this landscape, we examine how Community Notes behavior has changed, highlight its current limitations, and propose an integrative component built with the aim of accelerating the publication of notes. We first build a large-scale dataset of visual-based Community Notes. Using composition-adjusted regressions, discrete-time survival models, and a reconstruction of the Community Notes Core scorer, we highlight divergences in the current note life- cycle. The rating process for visual notes that reach Helpful status is slowing over time: in 2026, the rating phase takes, on average, 25.5% longer to publish than in 2024, which corresponds to an additional 2.1 hours. We show that this slowdown is associated with weaker rating accumulation, with a 30.7% decrease in the first 10 hours and increased overall system pressure. To counteract this evidenced slow-down, we propose a neural decision-support component that uses features available at successive rating stages to identify notes that can be safely published earlier. On our held-out test set, the proposed component can accelerate publication for 15.5 ± 2.6% of the helpful notes, reducing publication time by approximately 1.1 hours per note without introducing additional false positives.