Research Summary: Can We Still Trust Disaster Social Sensing? Empirical Evidence on Detecting AI-Generated Social Media Posts

A research study investigates the reliability of disaster social sensing in the context of generative Artificial Intelligence (AI) by examining whether text-based AI detectors can differentiate between human-authored and AI-generated social media posts. The study constructs a comprehensive dataset of 12,000 texts related to nine disasters, including human posts, AI-proofread human posts, factual AI-generated posts, and affectively framed AI posts, to test the efficacy of current detection methods. This research addresses a critical emerging challenge to information integrity in disaster response.

Why it matters

The increasing sophistication of generative AI poses a significant threat to the integrity of real-time information systems, particularly those reliant on public input like disaster social sensing. Organisations must understand the limitations of current detection technologies to safeguard decision-making processes, ensure accurate situational awareness, and maintain trust in information channels during critical events.

Key insights

• Generative AI can produce plausible social media messages resembling eyewitness reports, potentially compromising disaster social sensing.
• The study aims to empirically determine if existing text-based AI detectors can reliably distinguish human-authored content from AI-generated content in a disaster context.
• A dataset of 12,000 texts, comprising 3,000 matched semantic units, was created from nine distinct disaster events.
• The dataset includes original human posts (H0), minimally LLM-proofread human posts (H1), factual AI-generated posts (A0), and affectively framed AI-generated posts (A1).
• An additional 6,000-text corpus from 42 events was used for model selection and threshold calibration of AI detectors.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2609.35821