ResearchPod Summary
Public services, including policing, have historically relied on narrow, purpose-built AI systems to automate specific tasks like facial recognition or recidivism prediction. These systems were governed by established safety frameworks that focused on four pillars: accuracy, bias, explainability, and accountability. Because these tools were designed for a single, well-defined task, developers could measure error rates, audit training data for bias, provide feature-based explanations, and clearly delineate where human oversight began and ended. This created a manageable, evidence-based pipeline for safety assurance.
General-purpose AI (GPAI), however, represents a paradigm shift. Unlike narrow AI, GPAI models are trained on vast datasets to model the statistical structure of language, allowing them to perform an effectively unbounded range of tasks. This generality, combined with low deployment costs, has led to rapid adoption across public services. The authors argue that this shift renders traditional governance frameworks obsolete because the very properties that make GPAI useful—its flexibility and open-ended output—are the exact properties that make it impossible to govern using existing safety standards.
The paper details how the four core pillars of AI safety collapse when applied to GPAI:
The authors conclude that safety assurance has shifted from an intrinsic feature of AI development to an optional, external add-on, which is insufficient for high-stakes public services. They recommend a fundamental change in regulatory approach: establishing a clear taxonomic distinction between narrow and general-purpose AI, prioritizing technological parsimony, pausing operational deployment in policing until sufficient evidence exists, and creating a national safety infrastructure with the authority to independently verify the safety of these systems.
AI-generated third-party summary by ResearchPod. Not official content or an endorsement by the paper authors or affiliated organizations.