Why People Believe Misinformation — and What We Can Do About It

Key themes and claims

Historical perspective: The tobacco playbook

Disinformation follows a predictable playbook documented since the 1990s. When a 1994 whistleblower ("Mr. Butts") leaked ~4,000 documents to Stanton Glantz, they revealed three core tactics tobacco companies used to delay legislation:

1. Sow doubt — "Doubt is our product since it is the best means of competing with a body of fact." The goal is not to win with alternative facts but to establish controversy, rendering the consensus uncertain in the public mind.
2. Counterattack — "Attack the unprecedented nefarious attack against cigarettes." If you can distract from your own trouble and stay on offense rather than defense, that is a successful disinformation strategy.
3. Stick to facts — "Truth is our message because of its power to withstand conflict." While the playbook may include some lies, the most durable strategy relies on truthful or technically-accurate statements that are nevertheless misleading through framing and implication.

These tactics are not unique to tobacco. Roozenbeek observes them replicated in Russian Federation disinformation, climate denial campaigns, and fossil-fuel lobbying. The playbook recurs because the same strategic thinkers and PR professionals move between industries. Most critically, this historical perspective suggests that focusing on "fake news" (narrowly: false headlines) misses the most effective disinformation strategy: manipulation via selective truth.

Three theories of why people believe misinformation

1. Classical reasoning theory: People believe misinformation because they lack analytical thinking, cognitive reflection, or numeracy. Evidence exists (studies showing correlation between these traits and susceptibility), but Roozenbeek argues this tells only a small part of the story. Smarter people still believe false things; measured cognitive ability alone is not predictive enough.

2. Media consumption patterns: Exposure fuels belief. In echo chambers and polarized environments, repeated exposure to the same claims triggers the illusory truth effect — familiar claims seem more convincing over time. The more often you see a false or misleading claim, the more plausible it becomes, especially in homogeneous information diets.

3. Motivated reasoning and partisan bias: People preferentially agree with information their political side endorses, including misinformation. They also preferentially believe misinformation that makes the other side look bad. This is not about reasoning failure; it's about motivated reasoning — you believe what you want to believe.

Roozenbeek emphasizes the latter two as more important than classical reasoning theory, and draws several key insights:

• Persuading people of something is hard; exacerbating existing biases is easy. Making someone's pre-existing beliefs more extreme is much easier than shifting them from one position to another.
• Misinformation works better with longer and more intense exposure. The combination of duration and intensity matters; longer exposure plus greater frequency makes narratives more persuasive.
• The more persuasive the misinformation, the better for the propagandist. But if you can achieve your goals using truth rather than falsehood, that is preferable because it is harder to debunk. (E.g., the tobacco industry preferred to argue on real grounds about correlation vs. causation, or about alternative risk factors, rather than to simply lie. When disinformation actors must resort to outright lies—as Russia did with MH17 alternative theories—it signals they lacked a better option.)
• Disinformation is strategic, not ideological. A key claim: The observation that U.S. conservatives are more susceptible to and more likely to share misinformation is robust and real. However, this is temporary — it reflects that Russian disinformation currently targets right-wing populist groups as strategically useful (as they once targeted left-wing groups during the Soviet era). Framing this as "the right believes more misinformation" misses the deeper question: Why are right-wing groups being targeted more deliberately with misinformation right now?

Intervention landscape

Two levels exist for intervention:

System level — changing the platforms, incentives, and structures:

• Modifying recommender algorithms (limited effect; an Guess et al. study showed little benefit from reverse-chronological feeds)
• Business models and regulation (EU Digital Services Act, UK Online Safety Bill, U.S. has "absolutely nothing")
• Geopolitical pressure (attempting to persuade Putin to stop disinformation — Roozenbeek notes wryly this has not worked)

Individual level — targeting individual reasoning or behavior:

• Belief change via preemptive inoculation (teach people before they encounter misinformation) or postfact debunking/correction
• Behavior change via nudges, typically reducing the sharing of misinformation (not necessarily changing internal belief)

The efficacy-to-effectiveness gap: Six major challenges

Roozenbeek and collaborators identified six critical gaps between lab-study efficacy and real-world effectiveness:

1. Testing effects and the importance of rehearsal. A clever study showed that immediate testing after an inoculation video or game increases intervention longevity. A video-plus-immediate-test group retained effects 48 hours later; a video-only group (no test) showed complete decay of effects within 48 hours. Follow-up testing 10 days later showed similar patterns. Implication: People forget quickly without active rehearsal. For games, you can build testing in. For passive media (videos, articles), this is very difficult. Watching a video and moving on means rapid forgetting.

2. Lab-to-field attenuation. Lab studies show effect sizes (e.g., d = 0.40). When comparing to real-world outcomes, a meta-analysis by Stefano Deviatko and Elizabeth Lenos (Harvard) shows effects shrink by a factor of approximately 6. This means: Start with a robust lab effect, or it will vanish in the field. Many null findings in field studies may reflect small lab effects scaled down to zero.

3. Overestimation of intervention beneficiary populations. Researchers often compare treatment vs. control groups and claim an intervention works. But the actual beneficiary pool is much smaller than total population. For sharing-reduction interventions, the successfully-treated group consists only of people who:

• Are in an environment where an intervention can be deployed (e.g., on social media)
• Share unwanted content
• Share accidentally or out of ignorance (not maliciously)
• Are amendable to intervention
• See and interact with the intervention as intended (watch the whole video, play the full game)
• Encounter unwanted content before the intervention's effect decays
• Would have shared that content absent the intervention

This is a small subpopulation. It explains why field studies often show small or null effects — the denominator is not "all people" but a carefully-defined subset.

4. Lack of replication in the global south. Interventions validated in Western contexts (U.S., UK, Northern Europe) often fail to replicate in low- and middle-income countries. Reasons include underestimation of on-the-ground complexity, different media environments, cultural factors, and study-design issues. Roozenbeek emphasizes he is not an expert in this domain but notes the persistent pattern and its implications for equity and scalability.

5. Reactance and inoculation hesitancy. A paper by Alexander Johnson and Jens Madson (LSE) examines whether people will even engage with an intervention once they see it. Results show that engagement depends critically on the source of the intervention. No-source interventions see moderate engagement; high-trust sources (Harvard) maintain engagement. But Meta, the Republican Party, the Democratic Party, and the Russian government all show significant trust deficits, reducing engagement. This is called "inoculation hesitancy" — an analogy to vaccine hesitancy. Even if an intervention is effective, if people won't engage with it due to source distrust, effectiveness is moot.

6. Assumption that item-rating tasks predict real-world behavior. Lab studies measure success via "item-rating tasks" — participants see a tweet or headline and evaluate its reliability. We assume these task outcomes translate to real behavior (actually not sharing misinformation). But this assumption is largely untested. The few direct comparisons suggest the relationship holds but may be much weaker than assumed.

Solutions

Roozenbeek and colleagues propose multiple approaches:

Co-design with target audiences. Involve people from communities you're trying to reach in designing interventions; don't impose solutions from outside.

Invest in field studies. Labs are necessary but not sufficient; real-world testing is essential.

Elevate alternative outcome measures. Don't rely solely on item-rating task performance. Measure intervention attractiveness, engagement, sustained behavior change, and real-world sharing patterns.

Combine system-level and individual-level approaches. Individual interventions are insufficient if algorithms and platform incentives systematically promote manipulative content. Both levers must be pulled.

Disrupt the economics of manipulation. Roozenbeek highlights a massive, thriving market for fake followers, likes, comments. On Instagram (as of 2022), €10 buys 8,361 fake followers. This is absurdly cheap and enables artificial amplification of misinformation. Disrupting pricing and availability of these services could significantly impede manipulation campaigns. However, tracking price fluctuations and effectiveness of countermeasures has been difficult.

Case study: Strategic unfollowing on social media

Recent research by Steve Rache (NYU) and Roozenbeek demonstrates durable effects from a targeted individual intervention. Three studies (~3,000 participants, including correlational and two digital field experiments) tested what happens when people strategically unfollow specific polarizing accounts on X (formerly Twitter).

Hypothesis: It's not social media per se that drives polarization and misinformation exposure, but certain individuals within your feed. By unfollowing specifically polarizing accounts, you reduce both outgroup animosity and misinformation exposure without quitting social media entirely.

Results: Over 11 months (primary compliance window), people who unfollowed polarizing accounts and continued to do so saw improvements: their feeds felt more positive, intriguing, and educational; less sad, divisive, polarizing, upsetting, and angry. Critically, the effect was durable — a one-time intervention (unfollow once, then never again) persisted for ~a year.

Implication: This is an example of a system-plus-individual hybrid: individuals make a choice (unfollow), but the underlying system (feed composition) changes, and the net effect is substantial and persistent. This contrasts with other findings that simply deactivating social media entirely has limited effects on polarization.

Notes

This talk is Roozenbeek's most comprehensive synthesis to date of the field's theoretical foundations and practical limitations. Rather than simply advocating for interventions, he rigorously examines the gaps between what works in lab conditions and what works in the real world.

The tobacco playbook frame is particularly valuable. By grounding the discussion in documented, historical disinformation strategy, Roozenbeek sidesteps the risk of treating misinformation as chaotic or random. The playbook is deliberate, evolved, and replicable—and crucially, it relies heavily on selective truth and framing rather than outright lies. This legitimizes the CLAUDE.md principle that pages like this should focus on factuality over sensationalism: disinformation that clings to truth is more durable than disinformation that lies.

The prioritization of motivated reasoning over analytical ability is important for policy. If the problem were purely cognitive (people lack the ability to reason well), the solution would be education. But if the problem is motivated (people want to believe partisan-aligned narratives), education alone is insufficient—you must also change media environments, incentives, and social norms.

The honesty about intervention limitations is rare and welcome. Roozenbeek does not claim that inoculation games or videos are silver bullets. He carefully documents what they do help (recognition of manipulation techniques), what they don't (they don't change most people's beliefs, only those in a specific subpopulation), and what remains unsolved (reaching skeptical audiences, persistence across cultures, system-level causes). This intellectual maturity is the hallmark of good science.

The strategic unfollowing case study is instructive because it shows that even individual-level interventions can have system-level effects. By changing your feed composition, you change the content you see, which changes your downstream outcomes. This is simpler and more durable than trying to inoculate people against every false claim they encounter.

The inoculation hesitancy concept opens an important research gap: vaccine hesitancy has received significant attention; inoculation hesitancy (resistance to engaging with an intervention for reasons of source distrust or prior skepticism) has not. If Meta or political parties deploy anti-misinformation interventions, trust deficits will render them ineffective regardless of their psychological quality. This has profound implications for who should deploy interventions and how.