Think of a neighborhood watch, but for the health of democratic information. During the 2025 Australian federal election, a team of researchers from the University of Melbourne tested a bold idea: could trained volunteers, working from their own social media feeds, monitor and classify election-related misinformation in a way that is reliable, scalable, and genuinely useful? The project arrived at a moment of widespread concern about online falsehoods, especially in a country with compulsory voting, where voters have a legal responsibility to inform themselves. Professional fact-checking is rigorous but slow, while automated detection systems often struggle with the subtle, contextual, and culturally specific ways misinformation spreads. This study explored a middle path—citizen science. Over a six-week period from early April to mid-May 2025, 90 active volunteers collectively surfaced 592 potentially misleading posts and provided more than 3,000 separate assessments using a shared classification framework. They were not simply clicking “false” on content they disliked; they sorted posts into 17 categories, explained their reasoning, and rated their confidence. The result is an evidence-based glimpse of how ordinary people, with modest training, can act as distributed monitors of the information environment. And the findings suggest something more: that the act of monitoring itself may leave participants better equipped to tell truth from falsehood months later. This summary unpacks how the study worked, what it found, and why it matters for election monitoring and misinformation resilience.
How did the system actually work? The researchers recruited 419 people through targeted online advertising and professional networks, including a small group with expertise in misinformation, psychology, law, and communication. Everyone completed onboarding, but the active core was much smaller: 90 people who made at least one submission or assessment. These volunteers were asked to go about their normal social media use, but with an added layer of attention. When they encountered a post they believed was election-related misinformation, they submitted it through a custom-built web platform. For each submission, they chose a primary and optional secondary category from a detailed framework, wrote a written justification, rated how confident they were, and uploaded a screenshot. Then other participants could see the post and provide their own independent assessment, creating a peer-review loop. Before starting, volunteers completed training that included slides, a webinar on cognitive biases, and practical tips. They were explicitly told to set aside their political views and focus on whether content fit the shared definition of misinformation. This was important because the participant pool skewed left-of-center, reflecting the broader Australian electorate at the time. To reduce ideological filtering, the system required every assessment to include a written rationale, which past research suggests helps people evaluate content more carefully, regardless of their politics. The authors acknowledge that this cannot fully eliminate partisan bias, but the combination of training, justification, and peer review was designed to make the process as fair and transparent as possible.
What did the volunteers actually find? Over six weeks, they logged a combined 1,228 hours of engagement, with an estimated 614 hours spent actively monitoring their feeds. Contributions were unevenly distributed, as is common in volunteer projects: about 10 percent of participants accounted for 60 percent of submissions and 70 percent of assessments. Even so, the dataset was substantial. A thematic analysis revealed that a small number of recurring narratives dominated, and that these narratives shifted as the campaign progressed. Electoral integrity claims were present throughout, but they spiked sharply in the final weeks, becoming the most frequent theme around polling day. Themes related to social justice, immigration, and foreign threats also persisted throughout the campaign. The most common categories overall were false content, unproven or conspiratorial claims, and out-of-context material, which together accounted for nearly 60 percent of all submissions. When a target could be identified, the incumbent Labor government was the target in 53 percent of cases. The coverage was not limited to one platform: more than 90 percent of submissions came from just four major social media sites, with the remainder spread across YouTube, Bluesky, Reddit, Xiaohongshu, and Threads. To test whether this volunteer network produced coverage comparable to professional fact-checking, the researchers compared their dataset to 34 fact-check articles published by the Australian Associated Press during the same period. The themes matched closely: electoral integrity, government spending, and cost-of-living claims were prominent in both. Notably, every major election-related narrative that the fact-checkers deemed important also appeared in the volunteer dataset. The crowd did not capture everything, but it surfaced the same stories that professionals thought worth verifying, which is a strong signal that structured crowdsourcing can track significant misinformation narratives in real time.
The study didn’t just count posts; it asked whether volunteers could judge them reliably. After all, a monitoring system that produces chaotic, contradictory labels would not be very useful. Here, the results are reassuring, though imperfect. Participants did not treat every flagged post as misinformation. About 5.9 percent of submissions were classified as out-of-scope or not problematic, and 3.2 percent were coded as true or partly true. That shows evaluative judgment rather than an indiscriminate urge to label anything suspicious. When the researchers compared volunteer classifications to those of an independent assessor, the assessor agreed with the crowd’s classification in 81 percent of cases for posts with verifiable true/false status. A statistical measure of agreement, Cohen’s kappa, was 0.42, indicating moderate agreement beyond chance. That level is broadly consistent with other research showing that crowds can approach professional fact-checking accuracy under structured conditions. Among posts rated by multiple participants, mean agreement with the most common category was about 64.7 percent, though this may overstate independent agreement because participants could see each other’s classifications before submitting their own. Agreement varied by category: it was highest for categories with clear structural features, such as political scams, unproven or conspiratorial claims, and false content, and lower for more interpretive categories like rage-bait and cherry-picking. Confidence ratings followed a similar pattern—people felt more confident about clear-cut categories and less confident about ambiguous ones. Overall, the framework helped volunteers distinguish different forms of misinformation with meaningful consistency, while leaving room for ongoing discussion and disagreement.
The most intriguing result came three months after the election. The researchers wanted to know whether participating in structured monitoring was associated with any lasting change in people’s ability to distinguish true from false content. They conducted a follow-up assessment with 56 people from the active monitoring cohort and compared them to a newly recruited control group of 132 people who had not participated. Both groups completed a 20-item test made up of simulated social media posts with known truth values. The treatment group scored significantly higher, with an average accuracy of 73.8 percent compared to 67.5 percent for the control group. This difference corresponds to a moderate effect size, which in practical terms means the monitors were noticeably better at telling real from fake posts. The effect persisted even after adjusting for demographic differences between the groups, including age, employment, and education. But the authors are careful about what they claim from this result. Participants were not randomly assigned, so the difference could reflect pre-existing characteristics, such as stronger political interest, higher baseline media literacy, or greater motivation to spot misinformation, rather than any effect of the monitoring activity itself. Still, the result is consistent with a plausible psychological mechanism. Sustained, repeated engagement with misinformation—especially in an active, effortful way that involves justifying one’s judgments and seeing how others classify the same content—may act as a form of psychological inoculation. Instead of passively reading corrections, participants practiced the skill of critical evaluation in real time. That kind of practice is exactly what many media literacy interventions try to build, and it may be more durable than simple exposure to fact-checks. The authors describe the finding as associational rather than causal, but it points toward a promising avenue for future research.
What does this mean in practice? First, structured crowdsourcing offers a scalable and relatively low-cost model for monitoring misinformation during fast-moving events. Professional fact-checking cannot be everywhere at once, and automated systems still miss culturally specific and multimodal content. A trained volunteer network can fill some of those gaps, especially when participants are embedded in communities that centralized monitoring tends to overlook. In Australia, electoral misinformation targeting migrant and non-English-speaking communities, much of it circulating on platforms like WeChat and RedNote, is a recognized blind spot. Because those platforms often resist automated detection, community-embedded human monitors may be one of the few practical ways to see what is actually being shared within those spaces. The study did not test such a community-specific deployment, but the design clearly enables it. Second, the model is more than just a monitoring tool; it may also be a stealthy intervention. By asking people to log, categorize, and rate each other’s contributions, the system turns passive social media scrolling into an active exercise in critical thinking. It does not tell people what to believe—it gives them a framework for evaluating claims, and that skill may carry over to future content. This non-partisan, participatory approach has real appeal for educators, civil society organizations, and electoral commissions looking for ways to build resilience to misinformation without engaging in heavy-handed censorship. The study’s limitations are real: the active cohort was self-selected, ideologically skewed, and relatively small, and the follow-up result could reflect pre-existing differences rather than a true effect. But as a design evaluation, it demonstrates that ordinary people, with modest training and a clear framework, can do surprisingly rigorous work. In an age when misinformation spreads faster than fact-checks can keep up, enlisting the public as observers and evaluators may be one of the most promising ways to keep democracy healthy.