Key Contributions
- Accuracy nudges work at scale: A Twitter field experiment with 5,000+ users who had shared links from Breitbart, Infowars, or similar outlets found that a single DM asking "how accurate is this headline?" increased discernment 3-fold and improved average content quality by ~5%, with effects strongest on high-visibility accounts.
- Crowdsourced accuracy detection is viable: When 500 lay people rated headlines from Facebook's flagged content (without doing research, just gut-checking), their aggregate ratings correlated with professional fact checkers as well as fact checkers correlated with each other (~0.62 correlation at 15 lay raters per headline), suggesting wisdom-of-crowds can supplement expert fact-checking.
- Multiple light-touch interventions achieve similar effects: Asking people to rate accuracy, providing digital literacy tips, and priming the importance of accuracy all increased discernment 50–100% in lab experiments, with social-norm framings showing minimal benefit.
- Interventions improve decisions without broad behavior suppression: The nudges increase sharing of trustworthy outlets and decrease sharing of low-quality sites without suppressing engagement with other content types (e.g., personal posts, images).
Research Findings
Rand's Twitter experiment: Over 130,000 high-risk sharers were followed by bot accounts; 5,000+ of those who followed back received a neutral accuracy-rating prompt. The causal effect was measured by randomizing treatment timing. Result: a strong correlation between the outlets' fact-checker ratings and the change in users' subsequent sharing proportions—positive for mainstream news (Times, CNN, WaPo, Fox News) and negative for hyperpartisan/low-trust sites (Breitbart, Daily Caller).
Crowdsourcing headline quality: 209 headlines flagged by Facebook's algorithm were rated by 3 professional fact checkers and 500 crowdworkers. Correlation between lay-person consensus and fact-checker consensus: 0.62 (matching fact-checker inter-rater reliability), achieved with ~15 crowd ratings per headline.
Berinsky's intervention experiments: ~9,000 participants across five randomized studies tried various prompts: rating one headline, rating eight headlines with feedback, digital literacy tips, or social-norm messages. All significantly increased discernment (50–100% larger gap between sharing true vs. false content). Attentiveness checks confirmed effects hold for both all and high-attention participants.
Feed simulation: A realistic feed platform showed that both modal pop-ups ("Is this accurate?") and in-feed accuracy prompts increased sharing of true content while maintaining or decreasing false-content sharing.
Mechanisms
Berinsky and Rand argue that inattention to accuracy, not motivated reasoning, is the primary driver of misinformation sharing. Social-platform design (likes, followers, group signaling) makes people think about engagement, not truth. By making accuracy salient—without removing share buttons or imposing top-down truth judgments—users' own accuracy motivation activates, and they make better choices. Evidence: people are reasonably good at assessing accuracy when prompted, and effects persist across ideologies.
Implications for Platform Governance
Why not just fact-check everything? Fact-checking is bottlenecked by professional capacity and timing (stories go viral before labels appear). Platforms also risk the "flagging implies unflagged content is true" perverse effect.
Why not just block/remove content? Centralized gatekeeping raises concerns about political bias, legal liability (Section 230), and democratic legitimacy.
Why accuracy nudges? Users self-police based on their own accuracy judgment, no algorithmic truth arbiter needed. Nudges are lightweight, scalable, and preserve user agency. Business model: doesn't suppress engagement broadly (doesn't hurt dog photos, friend updates), only improves decision-making on news.
Crowdsourcing angle: Supplement fact-checkers with lay raters to scale coverage. The wisdom of crowds, aggregated properly, can feed ranking algorithms.
Wikipedia model (Tushman): Wikipedia's volunteer-driven, consensus-based fact-verification system is the world's fifth-most-visited website and demonstrates how a third-party, non-corporate entity can reliably adjudicate facts. Platforms could create a MediaWiki instance (limited to key topics: COVID, election integrity) to crowdsource fact-checking using Wikipedia processes, keeping existing business models and algos intact.
Limitations & Open Questions
- Persistence: Effects measured only up to 48 hours in the Twitter experiment; no long-term data.
- Generalization: All research on U.S. participants; applicability to other regions (religious/ethnic polarization, authoritarian contexts) unknown.
- Efficacy heterogeneity: Demographic variation in nudge effectiveness not fully explored (demographic analyses underway with Jigsaw collaboration).
- Crowdsourcing tradeoffs: NYU/Stanford companion study (referenced by Kelly) found tension between simplicity/transparency (aggregation) and accuracy (ML models); also between general crowds and politically sophisticated crowds.
- Political content vs. other domains: Most work focuses on news/politics; generalization to health misinformation, conspiracy, satire/parody edge cases requires further research.
Notes
This webinar (May/June 2020) captures a moment of optimism grounded in early experimental evidence. The framing differs from platform-centric ("What can Twitter/Facebook do?") and more toward behavioral science insight ("What can we activate in users?"). Tushman's emphasis on Wikipedia as a model is prescient given later 2020–2022 tech platforms' turn toward fact-checking partnerships; his open-source, limited-scope proposal avoids the "platform as arbiter" trap. The social-norm null finding (contrary to much behavioral-econ work) is notable—accuracy motivation may be more powerful than conformity pressure in this domain. Rand and Berinsky emphasize scalability, avoiding literacy-training burnout and reaching the most problematic sharers (high-engagement accounts, not just low-literacy ones).