Timing matters when correcting fake news

Contributions

• Directly compares the long-term efficacy of fact-checks delivered at three timing intervals: before, during, and after exposure, disambiguating competing theoretical predictions.
• Demonstrates that feedback mechanisms (debunking) produce substantially stronger one-week retention effects than inoculation-based (prebunking) or simultaneous-encoding approaches.
• Provides actionable design guidance for social media platforms: simple "true"/"false" tags applied post-exposure outperform real-time labeling of disputed content.
• Shows the effect persists across both experiments and applies to both true and false headlines, and does not depend on analytic thinking or political knowledge.

Method

Two nearly identical experiments (Experiment 1: N=1,362, Experiment 2: N=1,321) with identical design. Participants read 18 true headlines (from mainstream news) and 18 false headlines (identified as fabricated by Snopes.com) and rated their accuracy on a 1–4 scale.

Treatment conditions: Participants randomly assigned to one of four conditions:

• Prebunking: "True" or "false" tags appeared immediately before reading each headline (separate screen).
• Labeling: Tags appeared simultaneously during reading (on the same screen).
• Debunking: Tags appeared immediately after reading and rating (separate screen, following initial judgment).
• Control: No tags; participants rated headlines with no veracity information.

One week later, all participants re-rated the same 36 headlines for accuracy, this time without any tags or information.

Analysis: Final accuracy ratings from both experiments analyzed together via linear regression with robust standard errors, clustered on subject and headline. Model included dummies for treatment condition, headline veracity (false=0, true=1), study, and interactions between veracity and condition and veracity and study.

Results

False headlines (one week later):

• Debunking: β = −0.123 (F=16.34, P < 0.001) — significant reduction in false-headline belief
• Labeling: β = −0.081 (F=7.45, P = 0.006) — significant but smaller reduction
• Prebunking: β = 0.042 (F=1.74, P = 0.188) — no significant effect
• Debunking significantly more effective than labeling (F=25.39, P < 0.001) and prebunking (F=15.11, P < 0.001).

True headlines (one week later):

• All three treatments significantly increased true-headline belief vs. control (all F > 6.75, P < 0.01).
• Debunking significantly more effective than labeling (F=65.53, P < 0.001) and prebunking (F=47.02, P < 0.001).

Truth discernment (difference in accuracy judgments for true vs. false headlines):

• Debunking: 25.3% reduction in misclassification vs. control.
• Labeling: 8.6% reduction.
• Prebunking: 6.6% increase in misclassification (Experiment 1) or 5.7% reduction (Experiment 2).
• Effect robust across both experiments; no significant moderation by analytic thinking (Cognitive Reflection Test) or political knowledge despite both measures predicting baseline discernment.

Partisan alignment: Providing corrections after reading was somewhat less effective for headlines aligned with participants' partisanship than for non-aligned headlines (F=5.06, P = 0.025), but debunking remained significantly more effective than labeling or prebunking even for partisan-aligned headlines (P < 0.001 for all comparisons).

Notes

Strengths: The experiment cleanly isolates timing from other correction-delivery variables by holding the message content and valence constant across conditions. The one-week delayed test captures memory dynamics absent from immediate post-correction tests. Large N and replication across two experiments strengthen confidence in the effect. Analysis is preregistered and data/materials publicly available.

Limitations: The 1:1 ratio of true to false headlines does not reflect real-world social media (where true content substantially outnumbers false), so effects may underestimate the advantage of debunking when "false" tags are rarer and thus more surprising. Participants actively assessed accuracy at exposure, unlike real-world mindless scrolling, which might weaken initial impressions and reduce the advantage of feedback. The stimulus set is relatively small (18 true, 18 false), and while diversity of headlines was prioritized, generalization to other content domains is uncertain.

Mechanism: The results support the concurrent storage hypothesis — people retain both the false claim and its correction even after receiving it, but the correction fades over time. Debunking's advantage over prebunking is inconsistent with the simple "inoculation" account (where prior warning should prevent initial encoding of falsehoods). Instead, the data align with a feedback-based learning mechanism: initial judgment error followed by corrective feedback enhances long-term memory for the correction, even when the initial judgment was low-confidence. This resonates with research on the power of prediction error and surprise in boosting retention.

Practical implications: Social media platforms could improve the durability of fact-check impact by timing corrections post-exposure, either through delayed pushes or post-read summaries. However, this design must balance against the platform-design constraint that debunking is typically delayed (hours to days after false content spreads), whereas real-time labeling is easier to implement. The finding suggests that imperfect timing is better than immediate labeling for long-term truth discernment, but the practical trade-offs depend on platform economics and content velocity.