Too Little, Too Late: Moderation of Misinformation around the Russo-Ukrainian Conflict

TL;DR

This study investigates how misinformation about the Russo-Ukrainian conflict spreads on Twitter and how the platform's moderation responds. Using the AMUSED framework, it links 575 misinfo tweets identified by 41 Western fact-checkers to a large retweet sample (543,475 tweets) and enriches the data with geolocation and narrative annotations. The findings show that US-identified posters dominate misinfo activity, Ukrainian-origin accounts contribute significantly, and 84% of misinfo tweets remain accessible long after posting, while suspensions affect only a minority, highlighting a lag between spread and moderation. The work argues for design improvements in cross-language, narrative-aware moderation and better preservation of content history to enhance the integrity of online information during conflicts.

Abstract

In this study, we examine the role of Twitter as a first line of defense against misinformation by tracking the public engagement with, and the platforms response to, 500 tweets concerning the RussoUkrainian conflict which were identified as misinformation. Using a realtime sample of 543 475 of their retweets, we find that users who geolocate themselves in the U.S. both produce and consume the largest portion of misinformation, however accounts claiming to be in Ukraine are the second largest source. At the time of writing, 84% of these tweets were still available on the platform, especially those having an anti-Russia narrative. For those that did receive some sanctions, the retweeting rate has already stabilized, pointing to ineffectiveness of the measures to stem their spread. These findings point to the need for a change in the existing anti-misinformation system ecosystem. We propose several design and research guidelines for its possible improvement.

Figures (9)

• Figure 1: Pipeline diagram for data collection of misinfo tweets and a sample of their retweets.
• Figure 2: Daily number of misinfo tweets (in red, jittered vertically) and their retweets in our sample (black line). Four instances predating plot time frame excluded.
• Figure 3: Geographical distribution (map created with GeoPandas; boundaries are not exact and NaN is zero) of users (a) who have posted misinformation tweets and (b) those who have retweeted them. Colors are segmented by quintiles.
• Figure 4: Tweet flows between countries (smaller countries aggregated by continent, with "o." meaning "other").
• Figure 5: Number of retweets per misinfo tweet and the number of days until the last retweet.