Agentic AI
New AI model detects hazard information in social media posts to identify influence operations
Researchers developed a machine learning tool that identifies mentions of dangers and harms in text, revealing how disinformation campaigns weaponize hazard narratives during geopolitical crises.
A team of researchers has created a machine learning classifier to systematically identify when social media posts contain information about hazards, harms, or dangers. Unlike sentiment analysis or emotion detection, which capture emotional tone, this approach directly measures the presence of threat-related content—a signal that has been largely ignored in prior social media research.
Why hazard detection matters#
People naturally pay closer attention to negative information and threats, and statements about potential dangers tend to carry greater perceived credibility than claims about benefits. This psychological pattern makes hazard-focused messaging particularly effective for influencing public opinion. By building a model that isolates hazard mentions from general negativity, the researchers can better understand how information is being framed and shaped.
Testing on real-world conflicts#
The team trained their classifier on annotated posts from X (formerly Twitter) and then applied it to two major geopolitical events: the 2023 Israel-Hamas war and the 2022 French national election. In both datasets, hazard-related language appeared frequently, but the distribution was revealing. Accounts identified as part of information campaigns—particularly those backing weaker parties in a conflict—mentioned hazards at notably different rates than typical users.
The researchers found that inorganic accounts (likely bot networks or coordinated campaigns) discussing civilian harm references, potentially to generate international support. The frequency and framing of these hazard mentions deviated markedly from organic accounts, suggesting deliberate strategic messaging.
Advancing detection of information warfare#
These findings represent preliminary work toward understanding how hazard narratives function as tools within information warfare. By distinguishing hazard signals from emotion or outrage, researchers can now track whether threat-focused messaging is being artificially amplified or concentrated—a hallmark of coordinated influence operations. The approach opens new avenues for detecting manipulation during crises when hazard information is weaponized to shift perceptions and drive geopolitical outcomes.
