Ai harms tracked faster using social media and large language models
Move Fast and Mend Things: Keeping Up with Evolving AI Harms Using Social Media Commentary
Human-Computer InteractionComputers and Society
Summary
AI systems can cause new problems that are hard to predict or notice after they happen. The authors created a method that uses large language models to analyze millions of social media posts and find different types of AI-related harms. They made a detailed list of 12 major harm categories with subtypes, capturing issues experts sometimes miss, like new kinds of privacy violations. Their approach helps spot emerging problems sooner, which can improve how AI is safely managed.
What this means in practice
- •For ai safety teams: Monitor evolving AI-related problems by analyzing social media discussions using the published dataset and taxonomy to detect harms quickly and comprehensively.
- •For online community managers: Use the AI harm categories to identify and address emerging user concerns related to AI tools in social platforms for better user support and policy adjustments.
Authors
Jacqueline Rowe, Animesh Srivastava, Sai Teja Peddinti, Seliem El-Sayed, Nina Taft
Abstract
The rapid deployment of AI systems has created socio-technical, psychological, and operational harms that can elude ex-ante threat modelling and ex-post incident tracking. We introduce an LLM-assisted thematic analysis pipeline to dynamically detect, categorise, and track emerging AI harms from large-scale social media data. Applying it to 5.7 million Reddit post summaries over 18 months (01/2025 to 06/2026), we curate and release a dataset of 575,000 AI harm-related posts and a bottom-up AI harm taxonomy of 12 categories and 47 subnodes. The taxonomy reliably covers established expert-defined risks while surfacing granular harms that top-down frameworks overlook, such as distinct forms of AI privacy violations. Temporal analysis surfaces evolving user-centric harms, such as agentic privacy and security breaches, premature AI adoption in the workplace, and grief from AI companion discontinuation. Our pipeline shortens harm-detection timelines and hereby complements efforts towards more participatory and responsive AI governance.