General-purpose AI breaks the assumptions of existing policing and public-service governance: its open-ended outputs, accessibility, and persuasive fluency render accuracy, bias measurement, explainability, and accountability frameworks insufficient unless governance explicitly distinguishes GPAI and builds independent safety infrastructure.

Main Finding

The paper argues that the arrival of general-purpose AI (GPAI) — large language models that produce open-ended natural‑language outputs and can be redeployed by prompt alone — fundamentally undermines the safety assumptions and governance tools developed for narrow, purpose‑built AI. In policing (the case study) this mismatch creates a high risk that public‑service AI governance frameworks will fail: accuracy, bias, explainability, and accountability become far harder or impossible to guarantee using existing methods, turning safety from an intrinsic part of system design into an optional, external add‑on. The authors recommend halting operational GPAI deployment in policing until rigorous, independent evidence and a national safety infrastructure exist, preferring technological parsimony and a clear regulatory distinction between narrow AI and GPAI.

Key Points

• Distinct regimes: Narrow (task‑specific) AI enabled tractable safety practices (measurable accuracy, disaggregable bias, feature‑attribution explainability, clear lines of accountability). GPAI breaks those assumptions.
• Four core safety concepts invert under GPAI:
  • Accuracy: No fixed standard of correctness for open‑ended outputs; construct validity and measurement fail for many GPAI tasks.
  • Bias: Bias becomes a structural property absorbed from huge training corpora and is harder to detect or disaggregate in free‑text judgements.
  • Explainability: Fluent rationalizations produce the appearance of explanation without fidelity to model reasoning.
  • Accountability: Outputs that convincingly persuade can displace or erode human oversight; the line between tool and decision‑maker blurs.
• Common mitigations in policing (expert evaluation and human‑in‑the‑loop oversight) presuppose narrow‑AI properties and are insufficient when marginal deployment costs fall to prompt writing and local staff adoption.
• “Structural inversion”: GPAI’s accessibility and low marginal deployment cost shift safety assurance from being embedded in system design to being an optional afterthought, increasing risk and reducing oversight feasibility.
• Policy recommendations: explicitly distinguish narrow vs. general‑purpose AI in governance, prefer simple (parsimonious) technology choices, pause GPAI operational use in policing until evidence exists, and build a coordinated, independent national safety infrastructure with authority to evaluate and certify deployments.
• While policing is the detailed case, the authors argue these problems generalize across public services (health, social care, benefits, etc.) where stakes and distributional effects are large.

Data & Methods

• Type: Conceptual and policy analysis / scholarly preprint (not peer‑reviewed yet).
• Unit of analysis: The GPAI model class (rather than any one product or interface) to highlight properties that travel across applications.
• Methods:
  • Literature synthesis of AI safety, public‑sector governance guidance, and empirical findings on narrow AI (e.g., recidivism prediction, facial recognition).
  • Comparative argument tracing how safety methods developed for narrow AI rely on bounded task definitions and why those methods fail for GPAI.
  • Case‑study focus on policing to illustrate practical consequences, referencing legal cases, policy documents (EU AI Act, UK DSIT plans), benchmark studies, and recent model evaluations.
• No original empirical dataset or quantitative experiment; conclusions derive from theoretical analysis + synthesis of existing empirical and policy literature.

Implications for AI Economics

• Misleading cost‑benefit calculations: Standard ROI claims for GPAI in public services (efficiency and labor savings) understate the unmeasured safety, legal, reputational, and remediation costs that arise when accuracy and harms cannot be reliably measured ex ante.
• Negative externalities and market failure: Low marginal cost and broad accessibility create rapid, decentralized adoption (including informal rollouts) that can generate systemic harms not internalized by purchasers — suggesting a case for public intervention, regulation, or insurance mechanisms.
• Procurement and investment strategy:
  • Favor technological parsimony: cheaper, narrower systems with measurable performance may produce superior social returns when risk and monitoring costs are included.
  • Procurement should require independent evaluation, transparent evidence of task‑specific safety, and mechanisms to internalize liabilities (contractual, insurance, and warranties).
• Public spending priorities:
  • The authors’ call for a coordinated national safety infrastructure implies upfront public investment (testing labs, certification bodies, monitoring regimes) that produces public goods: standardized evaluation protocols, reproducible evidence, and regulatory capacity.
  • Such investment can reduce information asymmetries and transaction costs for procuring agencies, but it is itself costly and requires clear mandates and independence.
• Labor and substitution effects: While GPAI lowers the cost of automating a wide array of public‑sector tasks, the uncertain quality of outputs and supervision costs may reduce realized labor savings; miscalculated substitution can degrade service quality and increase long‑run expenditures (litigation, corrective interventions).
• Distributional impacts and social welfare: Hard‑to‑detect structural bias in GPAI can produce concentrated harms to vulnerable populations, increasing social costs and potentially requiring targeted mitigation spending. Economic evaluations must account for distributional risks, not just aggregate efficiency gains.
• Regulatory timing matters: A precautionary pause (as recommended for policing) affects adoption curves and market incentives; delaying deployment reallocates private and public investments toward safer, more measurable technologies but may also slow potentially beneficial innovations. Policymakers should weigh dynamic tradeoffs and consider sunset/conditional deployment rules tied to evidence thresholds.
• Role for economic instruments: Taxes, liability rules, procurement weights, and subsidies can be designed to internalize GPAI externalities and steer adoption toward applications where evaluation and accountability are tractable.

Note: This is a preprint by Sam Relins and Daniel Birks (University of Leeds) submitted for publication (not yet peer‑reviewed).