Professional and community-based fact-checking show different strengths, but neither performs strongly across trust, scalability, and impact

The paper, in plain language

The bet the platforms made

With Petter Bae Brandtzaeg, Silje Susanne Alvestad and Asbjørn Følstad, I systematically reviewed the 21 empirical studies published between 2022 and 2025 — the window in which community-based fact-checking went from experiment to operational governance on X and then Meta. We rated each study's evidence on three things a fact-checking system must deliver: user trust, scalability, and actual impact on misinformation beliefs and sharing.

Different strengths, shared weaknesses

Professional fact-checking earns more trust and measurable, if moderate, corrective impact — but it is slow and cannot match the volume. Community notes are fast and scale enormously, yet many notes never become visible because they fail cross-perspective consensus thresholds, often appear after a false claim has already travelled, and coverage skews with who participates. One comparative study found no significant difference in misinformation reduction between the two models despite the crowd's far greater throughput. Scalability at the point of annotation, we note in the paper, is not the same as scalability at the point of correction.

The honest print

The evidence base itself is lopsided: most studies concern X's Community Notes, more than half are North American, and most measure short-term belief or engagement rather than lasting behavioural change. Our three-way ratings were reached by consensus rather than formal reliability testing, because the studies were too heterogeneous to score on one rubric. And almost none of this research yet grapples with AI-generated misinformation, which will strain both models.

Why it matters

Fact-checking is not a courtroom for individual claims; it is epistemic infrastructure, and the current debate — professionals versus the crowd, censorship versus free speech — is too narrow to design it well. The evidence points toward hybrids: community speed for identification, professional review where reach and stakes are high. That conclusion connects directly to our Science paper on malicious AI swarms — as fabricated consensus becomes cheaper to manufacture, the systems that verify shared reality need to be designed for effectiveness, not just openness.