About the program
Brazil's Superior Electoral Court (TSE) treats coordinated disinformation as a first-order threat to the 2026 election. The operations documented in the 2024 municipal elections and the 2026 first round share a common structure: cloned candidate videos, counterfeit official-results livestreams monetized through Pix scams, coordinated rumor cascades, and synthetic narratives amplified by inauthentic engagement. They concentrate in the blackout window, when campaigns are legally barred from responding and false claims about results face no counter-speech.
The goal of this program is to close the measurement gap on the platforms where these operations are least visible to independent researchers. We combine lawful access to public content through the Meta Content Library and API with transparent, auditable detectors adapted from Vigia, our prototype monitoring system. The intended contribution is a validated detection method that civil-society monitors in other democracies can reuse.
Research questions
- Volume and timing. How does the volume and timing of election-related public content on Meta surfaces shift across the blackout window and runoff day, and which topics dominate each period?
- Cross-surface diffusion. How quickly do documented operation classes (counterfeit official broadcasts, coordinated rumor cascades, synthetic-candidate impersonation, paid political content during the blackout) move between live surfaces such as Twitch, Kick, and YouTube and Meta surfaces, and does diffusion accelerate during the blackout?
- Detection validity. What precision and recall do rule-based and model-assisted detection methods achieve against researcher-verified ground truth for each operation class, and where do they fail?
Methods
The study runs in three phases: instrument development (through mid-October 2026), live measurement during the blackout and runoff weekend (October 24–26, 2026), and validation and write-up (November 2026 through early 2027). All collection and analysis happens inside the Meta Secure Research Environment.
- Collection. Read-only Content Library API queries in Portuguese: keyword and hashtag panels (fraude, apuração, resultado oficial, candidate and institution names), target panels of public pages and groups documented by Brazilian fact-checkers and the TSE, and Ad Library queries for paid content. The study window runs September 1 through December 31, 2026, with October–November 2024 as a historical baseline from the last municipal election cycle. Retrieval stays within the 500,000-item weekly API budget.
- Detection. Transparent detectors: near-duplicate text clustering, coordinated-timing cascades, impersonation heuristics, and audience-anomaly signals. Flags from our external livestream monitoring (Vigia, covering Twitch, Kick, and YouTube) are used as seeds to find cross-posted copies on Meta surfaces. No generative models are used, and every verdict is inspectable.
- Ground truth. Stratified samples of flagged clusters are independently coded against a published codebook. We report precision and recall per operation class, with detection latency for the live window.
- Analysis. Volume curves by class and window, network analysis of cross-page content reuse, advertiser concentration, and detector evaluation.
Data protection & ethics
- Read-only charter. No posting, no account interaction, and no deanonymization of private individuals. Analysis is restricted to public content and platform-provided metadata.
- Enclave discipline. All queries execute inside the Meta Secure Research Environment, and only aggregate results leave the enclave. No raw post content is downloaded to local infrastructure.
- Human subjects. The study analyzes public content at the aggregate level and does not profile individuals. IRB guidance is being sought through UC Berkeley's Office for the Protection of Human Subjects.
- Legal context. Research use complies with Brazil's LGPD as applied to publicly available information for research purposes (Art. 7, XI; Art. 11, §2º, III).
- Outputs. Open working papers, a methods paper on detector validation, and aggregated artifacts shared with Brazilian fact-checking organizations (Aos Fatos, Comprova, Lupa).
Team
Department of Industrial Engineering and Operations Research (IEOR), College of Engineering, University of California, Berkeley. His work applies operations-research and machine-learning methods to detection problems in large-scale information systems. He built Vigia, the prototype monitoring system this program extends.
Collaborators may be added through the Meta research program invitation process.
Contact
ehovagim20@berkeley.edu
University of California, Berkeley
For correspondence regarding this research program.