Contributions
- Identifies and formalizes the implied truth effect: warnings on a subset of false content reduce belief in tagged items but increase belief in untagged items via comparative reasoning.
- Develops a Bayesian belief-updating model that predicts when the implied truth effect should dominate, including conditions under which it is eliminated (e.g., when untagged headlines are explicitly marked as unchecked).
- Demonstrates the effect generalizes from accuracy judgments to willingness to share headlines on social media (Study 2).
- Shows the implied truth effect is larger for politically concordant false headlines—a critical limitation for fact-checking policies.
Method
Study 1 (n = 5,271 MTurk participants): Participants rated accuracy of a mixed set of true and false political headlines (from 2016-2017). Three conditions: (1) control (no warnings), (2) warning treatment (6 false headlines flagged "Disputed by 3rd Party Fact-Checkers," 6 untagged), (3) warning + verification (some true headlines labeled "TRUE"). Participants rated each headline on a 4-point accuracy scale and reported willingness to share on social media.
Study 2 (n = 1,568 MTurk participants): Similar design with more explicit conditions—control, warning (12 false headlines tagged "FALSE," 12 true untagged), and warning + verification (12 false tagged "FALSE," 12 true tagged "TRUE"). Stronger, more obvious warning language. Measured both accuracy judgments and sharing intentions.
Both studies included partisan counterbalancing (pro-Democrat/pro-Republican headlines matched for intensity) and analyzed headline veracity, political concordance, and treatment interactions using logistic and linear regression.
Results
Study 1:
- Warnings reduced perceived accuracy of tagged false headlines (M = 0.187) vs. control (M = 0.220), p = 0.001.
- Untagged false headlines in the warning condition were rated more accurate than in control (M = 0.239 vs. M = 0.213), implying a comparative truth effect, p = 0.001.
- Warning effect was roughly twice as large for politically concordant headlines (p = 0.003).
- Implied truth effect was not eliminated by adding verification tags ("TRUE") to true headlines.
Study 2:
- Warnings reduced sharing of tagged false headlines (16.1%) vs. control (29.8%), p < 0.001.
- Untagged false headlines in warning condition were shared more often than control (31.0% vs. 26.0%).
- Implied truth effect on sharing was significantly larger than on accuracy judgments.
- Verification tags (marking untagged items as "unchecked") eliminated the implied truth effect—a key theoretical prediction confirmed.
- Political concordance again moderated the warning effect; the effect was largest for politically concordant false headlines in the warning + verification condition.
Notes
The implied truth effect is a significant constraint on deploying warning labels as a misinformation-control strategy. The finding that comparative reasoning induces a truthfulness halo on untagged content has direct policy implications: if only a subset of false headlines can be tagged (a realistic constraint), users may infer that untagged content is true. The paper's Bayesian model offers a principled explanation for this: in the absence of explicit verification (Study 2 shows this resolves it), people use tagging status itself as a signal of veracity.
The effect is strongest for politically concordant headlines, which aligns with motivated-reasoning theories but suggests the backfire is not driven purely by motivated skepticism—it appears as a general inference mechanism. The study's use of Mechanical Turk limits generalizability, though the effect sizes are consistent across studies and conditions.
This work has direct implications for social media platforms designing content-warning systems and fact-checking organizations attempting to scale their efforts. The key actionable finding: ensure that warning systems are comprehensive and make the "unchecked" status of unwarned content explicit to avoid the appearance of a truth endorsement.