As an AI Language Model, "Yes I Would Recommend Calling the Police": Norm Inconsistency in LLM Decision-Making

LLMs make inconsistent decisions about calling police: The study found that GPT-4, Gemini 1.0, and Claude 3 Sonnet often recommended calling police even when no actual criminal activity was present in home surveillance videos, and conversely, sometimes failed to recommend it when criminal activity was occurring—suggesting these AI systems lack reliable judgment for high-stakes safety decisions.

Neighborhood racial demographics influence AI recommendations: The models showed bias based on the racial composition of neighborhoods in the videos, meaning an identical situation might trigger a police recommendation in one neighborhood but not another depending on its demographic characteristics—a concerning finding for an application that directly affects public safety and community policing patterns.

Current bias detection methods are insufficient: The researchers found that existing strategies for detecting and reducing bias in AI systems failed to prevent norm inconsistencies in this normative decision-making context, suggesting that companies deploying AI for surveillance and safety decisions may have false confidence in their bias mitigation efforts.

AI arbitrariness poses real-world risks: Because these inconsistencies occur in the specific context of police recommendations tied to home surveillance, the arbitrary nature of LLM decisions could lead to discriminatory policing outcomes and unequal treatment across different communities and demographics.