Reducing misinformation sharing on social media using digital ads

The misinformation problem and existing interventions

Current state of art: Platforms use two primary content-specific approaches:

1. Demotion / algorithmic downranking — flagged misinformation is deprioritized in feeds
2. Warning labels — fact-checker labels reduce belief and sharing by 25–50%, with no evidence of ideological backfire

Both are effective per-item but don't scale: too much content, too slow to fact-check, and misinformation spreads before labels are applied.

The gap: These approaches require identifying specific false claims beforehand. What about content-neutral interventions that work regardless of topic?

Behavioral foundation: accuracy vs. attention

Rand's premise: most people do not want to share misinformation, but the social media environment is distracting. Users scroll through news mixed with cat videos and jokes, with social signals (likes, shares) highlighting engagement rather than accuracy. Even if someone could distinguish true from false if they thought about it, they simply forget to consider accuracy before sharing.

Implication: Getting people to think about accuracy—without providing fact-checks or telling them what to believe—should reduce misinformation sharing.

Prior survey experiment evidence

Over several years, Rand's group ran 20+ survey experiments (N > 26,000) with varied prompts:

• Asking participants to rate a random headline's accuracy before a study (priming accuracy)
• Simple tips: "Think carefully before sharing"
• Asking importance of sharing accurate news
• Telling participants others think accuracy is important

Meta-analytic result: Accuracy prompts reduced false claim sharing by ~10% with no effect on true claim sharing. This worked for:

• Political headlines (aligned and opposed to participants' views)
• COVID-related claims
• Across partisan groups (conservatives and liberals)
• Cross-culturally (16 countries, 6 continents): ~9.4% reduction

The intervention is content-neutral: no need to know in advance which specific claims are false.

Facebook field experiment

Design: Meta researchers partnered with Rand. Sample: 33 million Facebook users (upsampled 350,000 known recent sharers of misinformation links). Users randomly assigned to control (regular ads) or treatment (accuracy prompt ads).

Treatment: Over 3 weeks, treated users saw ~3 accuracy ads (replaced normal ads), each with prompts like:

• "Two tips for thinking critically before sharing an article"
• Short video: "Check for accuracy"
• Question: "How important is it to only share accurate news?"

Outcome measurement: Posts flagged as misinformation by Meta's classifier (combining third-party fact-checkers, community review, and algorithmic signals).

Results:

• Within browsing session (60 minutes after ad): 1.8% reduction in probability of sharing misinformation for all users
• Among prior sharers: 2.6% reduction (higher base rate of misinformation sharing made effect more visible)
• No effect on non-misinformation sharing: treatment did not suppress all sharing—only misinformation
• No effect over full 3-week campaign: effect decays between prompts (attention is transient)
• Moderation analysis: no significant differences by age, gender, education; if anything, slight benefit for older users
• No ideological backfire: works across partisan divides

Adjusting for exposure: Under reasonable assumptions (~10–20% of users exposed to misinformation during session), treatment effect on exposed users: ~10–25% reduction, bringing it closer to lab effect size.

Twitter field experiments

Design: Rand's MIT group (led by postdoc House Lynn) ran three experiments (8–22 days) targeting active users who had recently shared links to low-quality news sites.

Treatment: Unlimited ad budget showing users an average of 3 accuracy prompts per day. Instead of repeating the same ad (which risks "banner blindness"), the campaign rotated through 50 different creatives, all prompting accuracy:

• Quiz: "Do you think this is accurate?"
• Tips: "Check the source, read the article"
• Attention-grabbing: varied visual styles
• University-branded (University of Regina, where co-author was located)

Unlike Facebook, prompts were not Twitter-branded, reducing platform association.

Outcome: Retweets of links to low-quality domains (below 70/100 quality score using aggregated expert ratings; see Houses et al. PNAS Nexus 2023).

Results:

• Intent-to-treat (all assigned users): 3.7% reduction in low-quality link sharing
• Treatment on treated (adjusted for ~60% ad delivery rate): 6.3% reduction
• No learning to ignore: no interaction with campaign day—effect sustained over 8–22 days despite repeated prompts
• Consistent targeting: works best on chronic sharers; no effect on users with zero prior low-quality sharing
• No unintended effects: no increase in sharing of non-low-quality content
• Cross-platform consistency: convergent evidence with Facebook

Subset experiment: Users who hadn't shared low-quality links at all showed no effect, confirming intervention targets vulnerable populations efficiently.

Effect sizes and comparison to literature

Lab vs. field magnitude:

• Lab experiments: ~10% reduction in stated sharing intentions for clearly false claims
• Field experiments: 2.6–6.3% reduction in actual sharing of flagged misinformation (broader category)

Why the difference is reasonable:

1. Treatment dilution: Many exposed people don't read ads carefully
2. Opportunity structure: Many users not exposed to misinformation in the window, so no opportunity to be treated
3. Measurement: Lab uses known false/true; field uses classifiers (imperfect but better than domain-level heuristics)

Comparison to other ad effects:

• Facebook ad meta-analysis (600+ experiments): 5% increase in purchase behavior
• COVID vaccine PSA campaigns (Facebook/Instagram, 800+ campaigns): ~1% improvement in vaccine attitudes
• Political advertising (ACRONYM, 2020 election): 0.3–0.4 percentage point turnout shift

Rand's 3–6% effect on sharing is in a credible range and compares favorably given the difficulty of the target behavior.

Network amplification

Individual-level effects are amplified system-wide through two mechanisms:

1. Cascade suppression: If one user doesn't share, their followers don't see it, so none of them share it downstream
2. Algorithm feedback: Reduced sharing signals to the platform that content is less valuable, reducing algorithmic ranking and reach

These multiplicative effects mean individual-level reductions translate to larger platform-level impact on misinformation spread.

Mechanism and scalability

Why ads? Not because ads are ideal (people ignore ads), but because they leverage existing infrastructure without requiring platform engineering. The same approach could be more effective if implemented as:

• Interstitial prompts (must click through on login)
• Banner notices (harder to ignore than ads)
• Native friction (required brief pause before sharing)

Content-neutral design: Because the intervention doesn't specify which claims are false, it:

• Works across topics, domains, and claim types
• Doesn't require fact-checking capacity at platform scale
• Avoids paternalism of deciding truth for users (empowers users' own accuracy judgments)
• Scales without adding fact-checkers or classifiers

Requires targeting: Efficiency depends on targeting people likely to share misinformation (chronic sharers, or during crisis windows), not broadcasting to all users.

Requires regularity: Single exposure wears off; need sustained, varied reminders (hence 50 different Twitter creatives).

Key findings across experiments

• Robust: works on Facebook and Twitter, across US and 16 countries
• Bipartisan: no interaction with political affiliation
• Specific: affects misinformation, not general sharing
• Scalable: no need to identify specific false claims in advance
• Persistent: no evidence people learn to ignore varied prompts
• Targeted: most efficient when aimed at frequent sharers

Related interventions and complementarity

Rand contextualizes accuracy prompts within a broader toolkit:

Content-specific approaches (highly effective but limited scale):

• Warning labels: 25–50% reduction but require prior identification
• Demotion: immediate and powerful but need to know what to demote

Content-neutral approaches (scalable but modest individual effect):

• Accuracy prompts: 3–10% reduction, no identification needed
• Pre-bunking (teaching about false claims in advance): large effects but requires knowledge of what false claims will emerge
• Media literacy training: mixed evidence, risk of inducing distrust rather than discernment

System-level approaches:

• Improving misinformation detection (faster, better classifiers)
• Crowd-sourced fact-checking (Twitter's Birdwatch/Community Notes)
• Incentive structures for responsible sharing

Rand argues content-neutral interventions fill gaps left by content-specific approaches: they work while identification systems catch up, and they help people use their own accuracy judgments.

Limitations and open questions

• Effect size varies: lab (~10%) > field (~3–6%), with multiple reasons (dilution, opportunity, measurement)
• Decay without reinforcement: Facebook 3-week campaign showed no overall effect—needs regular prompts
• Targeting necessary: inefficient to deploy to all users; works best on chronic sharers
• Still not a silver bullet: combined with demotion, labeling, and other approaches, this is one tool among many
• Unknown optimal implementation: how much variation, what frequency, what friction level maximizes effect while minimizing user annoyance?

Notes

This work exemplifies a shift in misinformation research toward behavioral and cognitive interventions that don't rely on determining truth. Rather than platforms being arbiters of fact, the approach acknowledges users' own (often good) judgment when attention is primed toward accuracy. It's particularly valuable because it's:

1. Implementable without major engineering: uses existing ad systems
2. Defensible politically: doesn't impose fact-checker judgments
3. Empirically validated: lab + two platforms + cross-cultural evidence
4. Synergistic: complements rather than replaces content-specific interventions

The finding that polarized, motivated users can be valuable contributors to fact-checking systems (see Rand's other work on Community Notes) and that simple attention prompts help those users exercise good judgment suggests the problem is not that people are fundamentally irrational, but that the social media environment deprioritizes rational evaluation of truth. Makings accuracy salient—"remember to think about whether this is accurate"—harnesses people's existing values and judgment.