Signal Paper argues public-sector AI governance frameworks break down under general-purpose AI, using policing as a case study

Summary

A paper argues that general-purpose AI (GPAI), systems built on large language models that can be directed by a prompt alone to perform an effectively unbounded range of tasks, undermines the conditions under which public-sector AI safety has historically been achieved. It contends that the safety concepts foregrounded by public-service governance frameworks, such as accuracy, bias, explainability, and accountability, were made tractable by narrow, purpose-built AI, and that the mitigations prescribed in guidance documents presuppose exactly what GPAI removes. Using the case of policing, where the authors say governance failures carry the most severe consequences, they argue accuracy cannot be quantified over unbounded outputs, bias cannot be disaggregated when outputs are free-text judgments rather than categorical predictions, explainability gives way to a mere appearance of explanation, and accountability erodes as outputs are optimized to persuade. The paper contends that the two mitigations dominating current policing AI strategy, expert evaluation and human-in-the-loop oversight, both rest on assumptions that GPAI violates. The authors recommend a clear taxonomic distinction between narrow and general-purpose AI in governance documentation, a preference for technological parsimony, a pause on operational GPAI deployment in policing until adequate evidence exists, and a coordinated national safety infrastructure with authority to generate that evidence.