Understanding and Reducing the Spread of Misinformation Online

Key contributions

• Implied truth effect: Fact-checking labels work on flagged false content (50% reduction in sharing), but backfire by making unflagged false content seem more credible—a critical limitation for scaling fact-checking interventions.
• Source prominence fails: Emphasizing publisher source via logo/banner has no effect on belief or sharing because most viral content already has plausible headlines from trusted or unvetted sources; source information adds no discriminative information.
• The accuracy-distraction hypothesis: People claim accuracy matters most when making sharing decisions, yet share false content roughly as often as true content. The disconnect reflects cognitive constraints—social media's design (likes, followers, reinforcement) shifts attention toward engagement and social signals, away from accuracy. People can distinguish truth when forced to focus on it.
• Accuracy priming: A one-shot intervention (asking subjects to rate one neutral headline for accuracy before the main task) reduces false-content sharing by 20% and triples the discernment gap (true vs. false). Works across political ideology, replicates across different headlines and representative samples, and does not occur for other primes (e.g., rating humor).
• Field validation on Twitter: A stepped-wedge experiment with 5,400+ known Breitbart/Infowars sharers showed that a direct message asking about accuracy doubled the shift toward mainstream sources (NYT, WaPo, CNN, Fox News) and away from hyper-partisan outlets (Breitbart, Daily Caller, Daily Wire). Effect size suggests accuracy interventions are more practical than platform-scale fact-checking.

Structure and findings

Problem framing

Misinformation is not binary (true/false) but a continuum. Two types matter: fake news (entirely fabricated stories) and hyper-partisan content (biased or misleading coverage of real events). Hyper-partisan content likely has greater reach and perniciousness because it's less obviously false.

Fact-checking labels: Limited and backfiring

Experiment 1: Subjects shown a mix of true and false headlines with 75% of false ones labeled "disputed by fact-checkers":

• Labeled false content: 50% drop in sharing (strong effect).
• Unlabeled false content: significant increase in sharing relative to unlabeled true content.
• The "implied truth effect"—absence of a label implies checking and validation—is ~33% the size of the warning effect.

Implication: If fact-checkers miss >33% of false content, the intervention net-increases belief in false claims.

Source-prominence interventions: No effect

Experiment 2: Adding a large publisher-source banner to headlines:

• No effect on accuracy judgments or sharing, even in rigorous conditions.
• Root cause: Most viral headlines cluster along a diagonal in belief-space: headline plausibility and source trustworthiness align, so source information is redundant.
• Manual fact-check of low-trust outlets: headlines they publish are predominantly true; emphasizing the source would mostly suppress true content, not false.

Conclusion: Source cues don't add discriminative information for most content because outlets (trusted or untrusted) produce mixed-quality content.

Cognitive mechanism: Can people tell truth from falsehood?

Experiment 3: Subjects rate accuracy vs. willingness to share:

• Accuracy ratings: True stories rated vastly more accurate than false ones; partisan bias is small.
• Sharing behavior: No significant difference between true and false; people twice as likely to share false partisan-aligned content as to believe it's true.

Interpretation: The problem is not inability to discern; it's motivation. Social media context (likes, followers, social validation, ideological reinforcement) makes accuracy less salient. People have the competence but lack the incentive or cognitive bandwidth to prioritize accuracy in the social-media environment.

The accuracy-priming intervention

Experiment 4 (survey): Before the main sharing task, subjects rate one politically neutral headline for accuracy:

• Control: 27% sharing of false headlines.
• Accuracy prime: 21% sharing of false headlines (−20% relative reduction).
• True-headline sharing unchanged.
• Effect replicates across ideologies, samples (MTurk and Lucid), and new headline sets.
• Lumosity control: Rating headlines for "humor" instead does not reduce false sharing.

Mechanism: Priming activates the accuracy concept, shifting the attentional spotlight toward a criterion subjects already value. No need to persuade; just remind.

Field experiment: Twitter direct messages

Experiment 5 (stepped-wedge design, N=5,400+ Breitbart/Infowars sharers):

• Treatment: A direct message asking subjects to rate a neutral headline for accuracy; no persuasion or explicit intervention language.
• Outcome: Sharing in the 24 hours post-DM vs. before.
• Quality metric: Fraction of tweets to mainstream (NYT, WaPo, CNN, Fox) vs. hyper-partisan (Breitbart, Daily Caller, Daily Wire) and fake-news outlets, weighted by fact-checker trustworthiness ratings.

Results:

• Single-DM treatment doubled the shift toward mainstream sources.
• Improvement persisted in the 24-hour window; effect size consistent with lab estimates (~20% reduction in false sharing).
• No evidence of reactance or ideological backlash.

Implications

1. Scalability vs. accuracy. Fact-checking labels don't scale and risk backfiring. Accuracy interventions are simpler, less resource-intensive, and work across political divides.
2. Platform design matters. Social-media features (quantified feedback, echo chambers) create an environment where accuracy is easy to forget. Interventions that re-center accuracy align with people's own values.
3. Field validation required. Survey experiments show promise, but real-world effects are essential. The Twitter field experiment demonstrates that subtle, non-coercive accuracy prompts work in the wild.

Notes

This talk is a synthesis of Rand's lab and field experiments on misinformation interventions, delivered at GBH Forum Network (Boston-area public broadcasting). The core insight—that accuracy priming is more practical and scalable than fact-checking—became the foundation for later work on large-scale platform interventions. The Twitter field experiment is a particularly valuable contribution because it moves beyond self-reported sharing intentions to real behavioral outcomes with a hard-to-reach population (heavy misinformation sharers).

The talk emphasizes the importance of understanding why people share false content before designing interventions. By framing the problem as an attention/motivation issue rather than a capability deficit, Rand et al. open a path to lightweight, non-adversarial solutions that align with what people already value.