From moderation to chaos: Meta’s fact-checking and the battle over truth and free speech

Summary

This paper empirically tests the censorship narrative behind Mark Zuckerberg’s January 2025 decision to dismantle Meta’s Third-Party Fact-Checking Program. Combining a content analysis of 2,053 debunking articles from 16 Meta-partner fact-checkers across eight European and Latin American countries with 30 expert interviews, Cazzamatta shows that fact-checkers never had takedown authority and that, post-verification, content removal occurs in only about 30% of cases—most flagged content stays online, often unlabeled. Fact-checkers themselves overwhelmingly favor transparency, labeling, and counter-speech over deletion, distinguishing “freedom of speech” from “freedom of reach,” and view Community Notes as an inadequate, manipulable substitute. The paper reframes Meta’s shift as a move from “preemptive” to “conflictual” cooperation, raising particular alarm for under-regulated Latin American contexts.

Key Contributions

• First empirical evidence on what actually happens to content after fact-checker verification on Meta, directly rebutting the censorship framing.
• Cross-regional comparative design covering the two regions Zuckerberg explicitly targeted (Europe and Latin America).
• Triangulates large-scale content analysis with practitioner interviews to align observed platform behavior with insider accounts.
• Documents fact-checkers’ normative positions on free speech, transparency, and moderation.
• Extends Rone’s preemptive-vs-conflictual cooperation framework to the dismantling of the 3PFC program.
• Provides grounded skepticism about Community Notes as a replacement in polarized, weakly regulated contexts.

Methods

Quantitative content analysis of 2,053 debunking articles published in 2022 by 16 Meta-partner organizations in Germany, Portugal, Spain, the UK, Argentina, Chile, Brazil, and Venezuela, sampled from 11,636 links. An inductive codebook (developed via a 500-article pre-test) captured post-verification outcomes (labels, removal, court orders, no action, residual footprint, source pressure, voluntary corrections), coded by eight native-language assistants with intercoder agreement of 77–93%. Correspondence analysis mapped country–outcome associations (85% variance). Thirty semi-structured interviews with fact-checkers (Aug 2024–Mar 2025) were analyzed inductively in NVivo following Tracy’s approach. Country selection drew on Hallin and Mancini’s typology and deliberative democracy indices.

Findings

• Correspondence analysis showed cross-national uniformity rather than country-specific moderation patterns.
• Removal rates ranged from 21% (Venezuela) to 45% (Portugal), clustering near 30%, with attribution to Meta vs. original spreader often unclear.
• Platform labeling varied widely: 0% in Argentina (Chequeado uses screenshots), 10% in Portugal, up to 56% in Brazil.
• In many cases no action was taken—false posts remained accessible and unlabeled, undermining censorship narratives.
• Fact-checker requests for corrections from public figures rarely succeeded.
• Fact-checkers articulated ten arguments against removal, prioritizing speech protection, distinguishing “freedom of reach,” and warning that opaque takedowns fuel conspiracy theories.
• European fact-checkers welcomed the DSA and EFCSN-led advocacy; Latin American ones expressed pessimism about regulation and funding.
• Concerns about coordinated vote manipulation of Community Notes were grounded in experiences on X.

Connections

This work speaks directly to research on the post-2025 collapse of professional fact-checking infrastructures and the rise of Community Notes as a replacement, connecting to Allen2025-ot, DeVerna2025-dl, and Renault2025-uh on crowd-sourced verification, and to Bak-Coleman2025-pm and Donovan2025-ws on the political economy of moderation rollbacks. Its framing of conflictual platform cooperation and oligarchic moderation resonates with Rieder2025-ju, Helmond2026-ll, and Starbird2025-jj on platform governance and the weaponization of censorship rhetoric, while its Latin American focus complements Rossini2026-jn and Inacio-da-Silva2026-zf. Broader debates on misinformation interventions and labeling link it to van-der-Linden2026-jt, Lewandowsky2026-ob, and Budak2024-ef.

Podcast

A research-radio episode discusses this paper: Listen