Systematic meta-analysis of research on AI tools to deal with misinformation on social media during natural and anthropogenic hazards and disasters

The paper investigates how artificial intelligence tools are used to address misinformation on social media in the context of natural and anthropogenic hazards and disasters. The central research question asks what kind of gatekeepers (news moderators) social media algorithms and users should be when dealing with hazard- and disaster-related misinformation. The study situates the problem within the rapid diffusion of misinformation in the digital age, noting its psychological, social, and operational impacts during crises such as COVID-19 and events like Hurricanes Irma and the Manchester Arena bombing. Given that social media can amplify or attenuate risk perceptions and shape responses during emergencies, the authors emphasize the importance of AI-based moderation and detection tools, while underscoring the need to align such tools with human rights and journalistic ethics. The work aims to map the state of research, identify gaps, and provide insights for a balanced approach between algorithmic recommendations and user agency in gatekeeping.

An initial set of 37 review papers was examined to establish the state of the art; 24 were excluded for not centrally covering the combined themes of hazards/disasters, misinformation, social media, and computer-aided methods, leaving 13 directly relevant reviews. Across these, all focused on COVID-19 or disease outbreaks, with no coverage of other natural or anthropogenic hazards. Several reviews examined AI tools and methods (e.g., comparisons of techniques pre/post pandemic, content and impact analyses, and bot-focused studies). Missing in prior reviews were broader research variables such as the spectrum of research objectives pursued in the field, the disciplinary areas involved, the types of hazards beyond disease outbreaks, and the geographical location of funding sponsors. The present meta-analysis was designed to fill these gaps by encompassing both natural and anthropogenic hazards and systematically coding objectives, research areas, hazard types, and sponsor locations.

Data sources and time frame: Abstracts were extracted from Web of Science and Scopus for studies published up to July 2022. Initial retrieval yielded 246 (Web of Science) and 422 (Scopus) abstracts (total 668). Search strategy: Boolean keyword queries were applied to abstracts. - Hazard-related terms: (disaster) OR (emergenc*) OR (hazard) OR (flood) OR (earthquake) OR (industrial accident) OR (terrorist attack*) OR (COVID) OR (pandemic) OR (wildfire) OR (Coronavir*). - Social media terms: (social media) OR (Twitter) OR (WhatsApp) OR (Facebook) OR (Instagram) OR (YouTube). - AI/method terms: (detect) OR (monitor) OR (prevent) OR (screen) OR (AI) OR (artificial intelligence). - Misinformation terms: (fake news) OR (misinformation). The final selection required abstracts to satisfy the combined queries using AND across the three thematic groups and OR within each group. Screening and selection (PRISMA 2020): - Records removed: duplicates (n=163) and review papers (n=37) prior to screening. - Abstracts screened: n=468; excluded n=205 for not centrally referring to hazards/disasters, misinformation, social media, or computer-aided methods. - Full-text eligibility assessed: n=289; excluded n=23 for the same reasons. - Included studies: n=266. Coding and variables: For each included paper, the following were coded: publication year, research area, hazard type, related topics (via keywords), study objectives (general and sub-objectives), and sponsor location (when funding was stated, n=90 papers). For sponsor location, countries/regions were recorded and grouped into ranges by count. Analytical methods: - Descriptive statistics for publication year, research area distribution, hazard types, and sponsor locations. - Keyword co-occurrence network using VOSviewer. Author and index keywords were cleaned by merging synonyms (e.g., “coronavirus”→“covid-19”; “lstm”→“long-short term memory”; “nlp”→“natural language processing”). Nodes required at least 5 co-occurrences; clusters required at least 5 nodes. - Flow (Sankey) representation of research rationale using Sankey-MATIC. Five general objectives and 21 sub-objectives were identified via manual screening. Scoring assigned 1 point when a study addressed a single objective/sub-objective and 0.5 points when multiple objectives/sub-objectives co-occurred; totals were converted to percentages for visualization. Context on AI techniques: The study references common ML/DL techniques used for misinformation classification, including supervised methods (Naive Bayes, SVM, logistic regression, ensemble methods like AdaBoost, XGBoost, random forests) and DL approaches (CNNs, LSTMs), as part of the background framing of computational methods prevalent in the corpus.

- Publication trends: Research output increased notably from 2020 (32 articles) with a peak in 2021 (148 publications). One study appeared in 2023 during the evaluation window. - Research areas (share of corpus): Computer Science 50.3%; Engineering 12.8%; Medicine 12.4%; Social Sciences 5.8%; Humanities and Communication 3.3%; Mathematics 5.9%; Psychology and Neuroscience 1.0%; Physics 2.6%; Environmental Sciences 0.7%; Chemistry 0.1%; Business, Management and Decision Sciences 3.0%; Biology 1.8%. Social sciences and humanities are markedly underrepresented relative to the human-centered nature of misinformation. - Hazard types: COVID-19 dominates (92.0%). Other categories include Multiple Hazards 3.5%; Disease Outbreaks (non-COVID) 2.0%; Hurricane 1.0%; Earthquake 1.0%; Floods 0.5%. - Objectives focus: Most studies target detecting misinformation (68% of general objectives), with classification approaches comprising 52% of sub-objectives. Sub-objectives also span multilingual detection, bot-focused debunking, dissemination pattern monitoring, and analyses of heuristic processes, among others. - Keyword network: 71 nodes across four clusters. The largest nodes include “COVID-19” and “social media.” Red and blue clusters mainly capture scope-related terms (e.g., “infodemic,” “public health,” “information dissemination,” “rumor detection,” “vaccine hesitancy,” “Twitter”), while yellow and green clusters concentrate AI/analytical methods (e.g., “machine learning,” “supervised learning,” “topic modeling,” “sentiment analysis”), consistent with Computer Science’s prominence. - Funding geography: A small number of countries fund most work. The United States is the most frequent funder; China, Spain, and Italy are in the 14–16 paper range; several EU countries are in the 11–13 range. Saudi Arabia and UAE appear around 6; Brazil, India, South Korea, Malaysia, and Qatar around 3–4; numerous others (e.g., Australia, Canada, UK, Japan, Mexico, Norway, Switzerland, Taiwan, Thailand) in the 1–2 range. Countries with high COVID-19 mortality (e.g., US, Italy, Spain) also show higher publication counts, though this relationship is not uniform (e.g., Brazil, Mexico, Slovakia have relatively few publications despite high mortality).

The findings indicate that the field has been shaped predominantly by the COVID-19 pandemic, both in volume and topical focus, with a strong skew toward detection-oriented AI research. This emphasis addresses an immediate need—identifying unreliable content rapidly during crises—but leaves broader questions about the roles of different gatekeepers underexplored. Specifically, the balance between algorithmic recommendations (e.g., automated moderation, ranking, and classification) and user agency (e.g., informed decision-making, critical engagement) is insufficiently theorized and empirically examined. The pronounced dominance of Computer Science and relative lack of contributions from social sciences, humanities, psychology, and decision sciences suggest that critical dimensions such as ethics, human rights, behavior, cognition, and governance may be under-integrated into AI tool design and evaluation. The geographical concentration of funding underscores disparities in research capacity and digitalization across countries; while high pandemic burden may motivate investment, it does not consistently translate into research output. Collectively, these results point to the need for interdisciplinary collaboration and policy frameworks that align AI-driven gatekeeping with human rights and journalistic standards, and to extend insights developed during the pandemic to other hazards (e.g., earthquakes, floods, hurricanes, industrial accidents, terror attacks).

This meta-analysis maps the research landscape of AI tools to combat misinformation on social media in the context of hazards and disasters, codifying objectives, disciplines, hazard coverage, and funding geographies across 266 studies. It highlights a surge of COVID-19-centric, Computer Science–led, detection-focused work and reveals gaps in social science and humanities engagement, in research on non-COVID hazards, and in reflections on the complementary roles of algorithms and users as gatekeepers. The study contributes a structured overview and identifies priority directions: broaden research beyond COVID-19 to other hazards; increase participation from social sciences, decision sciences, psychology, and humanities; articulate and evaluate frameworks for balancing algorithmic moderation with user choice, grounded in human rights and journalistic ethics; and support research capacity in less digitally competitive countries. Future work should also compare these patterns with national digitalization indicators (e.g., in industry and education) to understand structural drivers of research leadership and impact, and translate high-level principles into operational guidelines for rapidly evolving social media contexts.

- Database and timeframe constraints: Only Scopus and Web of Science were queried, up to July 2022, which may omit relevant studies from other sources or later periods. - Keyword-driven selection bias: The reliance on specific AI/method and social media terms (e.g., “detect,” “monitor,” “AI”) may underrepresent work from social sciences/humanities that use different terminology. - Abstract-based screening: Although many full texts were assessed, initial inclusion decisions depended on abstracts, which may incompletely reflect study scope or methods. - Funding data incompleteness: Only 90 of 266 papers reported funding, limiting the robustness of sponsor location analyses. - Inconsistencies and rounding in reported categories: Some percentages and sponsor counts are grouped into ranges and may not sum perfectly due to rounding or categorization. - Generalizability across hazards: The overwhelming focus on COVID-19 limits conclusions for other hazard types where information dynamics may differ.