Why Public Service AI Governance Frameworks Risk Failing in the Age of General-Purpose AI: Lessons from Policing

Traditional AI safety measures don't work for general-purpose AI: The paper argues that governance frameworks built around narrow AI systems—which focus on measuring accuracy, detecting bias, ensuring explainability, and establishing accountability—become ineffective with large language models. For example, "accuracy cannot be quantified over unbounded outputs" when AI can generate any type of response rather than choosing from predefined categories.

The accessibility of general-purpose AI creates new risks: What makes GPAI attractive to cash-strapped public services—its generality, ease of use through simple prompts, and low deployment cost—actually undermines safety. The paper warns that these properties remove the conditions under which traditional AI safety has been pursued, making it harder to control how the technology is used.

Current oversight strategies are insufficient: The two main safeguards in policing AI (expert evaluation and human-in-the-loop oversight) both assume humans can meaningfully review and correct AI outputs. However, the paper suggests GPAI violates these assumptions, particularly when outputs are persuasive free-text judgments rather than simple categorical predictions that are easier to verify.

The problem extends beyond policing to all public services: While the paper focuses on policing as a case study where governance failures have the most severe consequences, the authors argue "the same failure is likely to recur across other public services" that adopt general-purpose AI under similar resource pressures.

Governance frameworks need fundamental redesign: The paper implies that incremental improvements to existing AI oversight won't suffice; public service AI governance must be reconceived for an era of general-purpose models rather than purpose-built systems.