Research Summary: Move Fast and Mend Things: Keeping Up with Evolving AI Harms Using Social Media Commentary

Researchers have introduced an LLM-assisted thematic analysis pipeline designed to dynamically detect, categorise, and track emerging AI harms from large-scale social media commentary. Analysing 5.7 million Reddit summaries between January 2025 and June 2026, the study generated a dataset of 575,000 harm-related posts and a 12-category, 47-subnode taxonomy that surfaces granular risks—such as distinct AI privacy violations—frequently omitted by top-down governance frameworks.

Why it matters

Standard AI threat models frequently fail to anticipate the full breadth of operational, psychological, and socio-technical harms experienced by users post-deployment. Incorporating bottom-up surveillance of public commentary allows organisations to identify unforeseen vulnerabilities before they manifest as critical failures or regulatory infractions. Transitioning to continuous, dynamic risk discovery is essential for maintaining institutional resilience in rapidly evolving technological environments.

Key insights

• Ex-ante threat modelling and ex-post incident tracking often miss socio-technical, psychological, and operational AI harms generated by rapid deployment.
• An LLM-assisted thematic analysis pipeline evaluated 5.7 million social media post summaries across an 18-month period, curating a dataset of 575,000 AI harm-related posts.
• A bottom-up taxonomy featuring 12 categories and 47 subnodes was established, successfully capturing both established expert risks and overlooked granular harms.
• Temporal analysis of public commentary provides an empirical mechanism for tracking emerging AI risks as they develop in real time.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2610.09082