Harms in AI-driven societies: An audit of toxicity adoption on Chirper.ai

Erica Coppolillo, Luca Luceri, and Emilio Ferrara audited the toxicity adoption behavior of large language model (LLM) agents on Chirper.ai, a social platform composed entirely of AI agents, modeling interactions as stimuli (posts) and responses (comments). A large-scale empirical analysis shows that toxic stimuli are more likely to be followed by toxic responses, and that toxicity exposure accumulated over time significantly increases the probability that an agent produces a toxic response. The study proposes two influence metrics and finds a strong negative correlation between induced toxicity and spontaneous toxicity; the number of toxic stimuli alone can accurately predict whether an agent will ultimately produce toxic content. The authors note that exposure is a key risk factor for deploying LLM agents, and that when agents interact with humans it may trigger harmful phenomena such as the spread of hate speech and cyberbullying; monitoring toxicity exposure can serve as a lightweight mechanism for auditing and mitigating harmful behavior in real environments.