Exploring Lightweight Interventions at Posting Time to Reduce the Sharing of Misinformation on Social Media

Farnaz Jahanbakhsh; Amy X. Zhang; Adam Berinsky; Gordon Pennycook; David G. Rand; David R. Karger

doi:10.1145/3449092

In brief

Users share fewer false headlines when asked at posting time to assess accuracy or provide reasoning for why a claim is accurate. A developed taxonomy of reasons why people believe or disbelieve claims enables structured interventions that can be integrated into platforms as signals of credibility. The interventions reduce both false and true content sharing, with greater reductions in false content, and the effect persists across confidence levels.

Contributions

Developed a taxonomy of reasons why people believe or disbelieve news claims through participatory design with 337 Mechanical Turk participants, refined over three iterations.
Tested three lightweight behavioral nudges at the point of sharing: (1) checkboxes indicating headline accuracy, (2) tagged reasons from the taxonomy, and (3) free-text rationales.
Evaluated interventions with 1,668 participants and found that both accuracy assessment and reasoning nudges reduce false headline sharing by ~17% and reduce true content sharing by ~18%, with reasoning nudges having an even greater effect.
Demonstrated that structured reasons can be aggregated and presented as credibility signals (e.g., "27 friends think this post is true").

Method

The paper presents two main studies:

Study 1 (Taxonomy): Participants were shown news claims (with images, lead sentences, and sources) and asked whether each was accurate or inaccurate. In two iterations, they provided free-text rationales for their beliefs. A research team iteratively refined these rationales into a taxonomy of categories representing why claims are believed or disbelieved. The final taxonomy includes categories like "Accurate on the evidence," "Inaccurate by contrary knowledge," "Claim is misleading," and others organized by perceived reliability and consistency with past experience.

Study 2 (Nudge Study): Participants viewed headlines one at a time and were randomly assigned to one of four conditions: (1) control (no intervention), (2) accuracy assessment (checking a box to indicate if the headline is accurate), (3) structured reasoning (selecting from checklist reasons why a post is accurate), or (4) free-text reasoning (writing an explanation). Before sharing, participants indicated their sharing intentions on a 3-point scale. Accuracy assessment, confidence, and other covariates were measured.

Results

Key findings on sharing behavior:

Accuracy assessment nudge: Reduced sharing of headlines assessed as inaccurate. Participants were less likely to indicate they would share false headlines in treatment conditions (mean ~0.08 vs. control ~0.20 for false; mean ~0.28 vs. control ~0.32 for true).
Reasoning nudge: Had an even larger effect. Asking for reasoning lowered the ratio of shared false to true content by 17% for false and 18% for true, with the decrease in false content being substantially larger.
Confidence interaction: Interventions were most effective when participants were less confident in their assessments. The interventions reduced sharing of headlines about which they initially expressed low confidence.
Perceived accuracy: Share likelihood increases with average perceived accuracy of a headline across all conditions, suggesting that interventions help people differentiate between headlines they perceive as true vs. false.

Structured reasons performed comparably to free-text reasoning in reducing false headline sharing while offering the advantage of being aggregatable for presentation as credibility signals.

Notes

Strengths: The taxonomy development is rigorous and participatory, grounded in how users actually reason about truth and falsehood. The finding that structured reasoning can serve as aggregatable credibility signals is novel and has direct platform-design implications. The paper tests multiple intervention designs and measures both false and true content sharing, providing nuance on the cost-benefit tradeoff.

Limitations: Study conducted via Mechanical Turk with a largely U.S. population; generalizability to other platforms and geographies is uncertain. The hypothetical sharing scenario may not fully predict behavior in deployed systems. The paper notes a possible Hawthorne effect: being aware of being studied may change sharing behavior compared to natural platform use. Future work should validate these nudges in real deployed systems.

Open questions: How do these interventions perform when exposed to the full diversity of true and false claims, not a curated set? Do the effects persist over time or does habituation occur? How should platforms balance encouraging sharing of true content with discouraging false content, given that both decline?