Accountability in AI and Automated Decision-Making: From Explainability to Reviewable Administrative Governance

Review

Tzu-Yu Chiou challenges the tendency to equate accountability in public-sector automated decision-making (ADM) with model explainability. Drawing on Lorenz von Stein's account of administration as a mediator between state and society, the chapter locates the central risk in the public sector's dependence on private suppliers of data, models, compute, and the categories through which administrative problems are perceived. Even an intelligible decision may be institutionally unaccountable when the administration lacks authority or capacity to reconsider it. Conversely, the author argues, some technical opacity may be tolerable where genuinely independent review and enforceable remedies remain available. The argument concerns public administration, not all deployments of AI.

The proposed reviewability principle has four interdependent elements: an affected person's right to challenge an automated decision and obtain substantive human redetermination; independent access to audit training conditions, logic, and population-level effects without vendor gatekeeping; scheduled or triggered reassessment independent of individual complaints; and judicial or quasi-judicial power to suspend, modify, or withdraw deficient systems. Explainability is retained as evidence supporting these institutional duties rather than discarded. The building-inspection example makes the distinction concrete: a model card cannot by itself establish whether historical enforcement data reproduces unequal inspections or enable correction of that pattern.

The method is reconstructive legal and political theory, not an empirical test of technology-company capture. The chapter adapts Stein's theory of class capture to contemporary dependencies in administrative cognition, then applies that framework to Taiwan's Artificial Intelligence Basic Act and existing administrative procedure, procurement, oversight, and litigation institutions. Its useful institutional move is to locate AI accountability inside enforceable administrative law and procurement rights, rather than establishing a parallel system of disclosures and conformity assessments. In Taiwan, the proposed bridge includes meaningful notification and reconsideration, mandatory vendor-independent audit provisions, recurring oversight linked to the Control Yuan's functions, and findings usable by administrative courts.

The structural analogy identifies a real vulnerability without proving its prevalence or severity across agencies, vendors, and decision classes. Control of technical infrastructure does not invariably equal control over policy or adjudication. The author anticipates this objection by treating capture as a dependency risk rather than an allegation of vendor bad faith and accepts calibrated safeguards for high-stakes rather than all ADM. That is persuasive as an institutional risk model, but not yet a demonstration of causal capture. Its historical novelty lies chiefly in reconstructing Stein as a theory of administrative epistemic dependence; rights to appeal, audit, and remedy themselves have substantial prior administrative-law lineage.

Implementation would require sharper specification of who obtains audit access, what information vendors must preserve and disclose, which institution can order interim suspension, how adversarial challenges are funded, how trade secrets and sensitive personal data are handled, and how reassessment triggers are verified. Nor does a nominal human redetermination ensure independence if officials cannot interrogate a vendor's evidence or are incentivized to ratify its outputs. A useful next test would follow one high-impact procurement through contracting, model updates, contested decisions, independent audit, remedy, and withdrawal, measuring whether the public authority can act without vendor permission. The chapter establishes reviewability as the relevant constitutional and administrative design objective; the remaining work is to show that review powers can survive the infrastructure dependencies they are intended to discipline.

Key Insight

An automated public decision is accountable only when public institutions can independently examine its conditions of production and compel its correction, not merely explain its output.