When Users Turn Hostile: Rude, Aggressive, and Abusive Interactions with AI Chatbots

Abstract

AI chatbots are increasingly exposed to hostile user behavior, yet prior work often collapses such inputs into a single “toxicity” label. This study reframes hostility toward conversational AI as an interactionally structured phenomenon and introduces a pragmatically grounded annotation scheme that distinguishes rude, aggressive, and abusive behavior based on dominant communicative intent, responsibility attribution, and interactional function. Using an ethically compliant hybrid corpus of 1,200 user utterances across 180 conversations—combining controlled synthetic interaction breakdown scenarios with curated publicly reported interaction patterns for qualitative grounding rather than prevalence estimation—we conduct a multi-layered analysis integrating discourse interpretation, linguistic feature profiling, and auxiliary computational metrics. Inter-annotator reliability is substantial (Krippendorff’s α = 0.72), with the highest ambiguity observed at the rude–aggressive boundary. The results reveal that hostile behavior frequently escalates along coherent interactional trajectories linked to system breakdowns such as incorrect responses, refusals, and repetitive output: rudeness typically signals early dissatisfaction, aggression introduces confrontational blame and dominance cues, and abuse shifts toward expressive degradation and dehumanization. These findings demonstrate that monolithic toxicity modeling obscures interactionally meaningful distinctions and motivate severity-sensitive, context-aware response strategies for safer and more resilient conversational AI. The proposed framework offers a transferable foundation for severity-sensitive analysis and response design in conversational AI. This work was conducted at Arab International University (AIU), Damascus, Syria.Official website: https://www.aiu.edu.sy

Ähnliche Arbeiten