Does the Source of a Warning Matter? Examining the Effectiveness of Veracity Warning Labels Across Warners

TL;DR

This paper investigates whether the source of veracity warning labels affects trust in and sharing of false information on social media. Using a large $N=2049$ online, between-subjects experiment with five conditions (No label, Platform, Crowd, AI, Fact-checkers), it measures post-level trust and sharing across false and true posts and analyzes demographic and attitudinal moderators. The main finding is that all warning-label sources reduce trust in false information, with AI labels yielding the strongest effect, and, to a lesser extent, on sharing intentions; the effects are moderated by prior trust in media and by the information itself, and AI warnings are particularly effective for participants with low trust in news organizations. The study highlights both the potential and the caution required when deploying AI-based veracity labels and suggests that warning-label design should consider user trust profiles, while urging longitudinal and real-platform evaluations to establish real-world impact. Overall, the work contributes to understanding how source cues shape warning-label effectiveness and informs better design and testing of misinformation interventions in online ecosystems.

Abstract

In this study, we conducted an online, between-subjects experiment (N = 2,049) to better understand the impact of warning label sources on information trust and sharing intentions. Across four warners (the social media platform, other social media users, Artificial Intelligence (AI), and fact checkers), we found that all significantly decreased trust in false information relative to control, but warnings from AI were modestly more effective. All warners significantly decreased the sharing intentions of false information, except warnings from other social media users. AI was again the most effective. These results were moderated by prior trust in media and the information itself. Most noteworthy, we found that warning labels from AI were significantly more effective than all other warning labels for participants who reported a low trust in news organizations, while warnings from AI were no more effective than any other warning label for participants who reported a high trust in news organizations.

Figures (5)

• Figure 1: Example false post from each warning label condition. Both the Crowd and Platform conditions match the respective population (Facebook or Twitter). Note that each participant also saw four true posts with no labels in each the condition.
• Figure 2: Correlation of demographic variables across both platforms. Details are in the supplemental materials.
• Figure 3: Cohen's d effect sizes relative to control across conditions, where participant-level are effects between participant-level scores (-4 to 4) and post-level are effects between post-level scores (-1 to 1).
• Figure 4: Interaction plots of participant-level trust across high and low ABINews and ABISocial subgroups. All interactions were significant except ABINews X AI. Note, the y-axes are at different scales.
• Figure 5: Participant-level trust scores across conditions and self-reported political leaning. Conservatives includes participants who claimed to be either conservative or very conservative (a). Moderates includes participants who claimed to be moderate (b). Liberals includes participants who claimed to be either liberal or very liberal (c). In (d), we show the distributions across all participants.