Algorithmic Authority in Practice: Generative AI, Decision-Making, and Accountability in Public Administration

TL;DR

Drawing on 24 semi-structured interviews with Kenyan public officials, this interpretive study argues that generative AI reshapes public administration through three interrelated mechanisms: concentration of epistemic authority in fluent narrative outputs, default compliance with an asymmetric burden of justification for overrides, and diffusion of accountability through system-based attribution. The paper frames these as conditional rather than universal effects, and treats implications for citizen trust as officials' perceptions rather than direct evidence.

Key findings

• Epistemic authority concentrates in GenAI outputs, most strongly among officials lacking technical or situated knowledge; the paper attributes this to the persuasive form of fluent, full-sentence recommendations rather than to numerical scores or binary flags, though it states this is "theoretically significant rather than formally demonstrated."
• Decision behaviour follows a structured pattern of default compliance and selective override: deviation required written justification while adherence did not, so "the burden of proof" shifted onto human decision-makers. Overrides were reactive, triggered by conflict with local knowledge, and scepticism was experiential (arising from prior system error) rather than procedural.
• Accountability diffused via system attribution ("the system flagged it"); absence of clear accountability frameworks emerged spontaneously in 14 of 24 interviews, across all levels and all five organisations. One counterexample in a grants unit showed accountability remaining traceable where a named reviewer documented an override.
• Officials perceived that opacity and diffused responsibility could strain institutional trust even where performance improves; positive trust effects were described as contingent on transparency and comprehensible outcomes.

Methods

• Interpretive qualitative embedded single-case design: five purposively selected Kenyan public-sector organisations (anonymised A–E) participating in the national RCoC/KSG–MICDE–UNDP–Microsoft AI skilling initiative, spanning central policy, agency operations, regulatory oversight, statistics/planning, digital services, and frontline provision.
• 24 semi-structured interviews conducted November 2025–March 2026 (11 face-to-face, 13 online), each 55–70 minutes; participants had 3–18 months of GenAI use and all used it at least weekly; 2 were involved in system design/fine-tuning, 22 were end users. Systems were commercial LLMs behind custom government interfaces, deployed 10 months to 2 years, advisory only with human sign-off.
• Thematic analysis following Braun and Clarke's six phases under an abductive logic; open coding produced 347 descriptive codes grouped under three sensitising dimensions (technocratic power, decision dynamics, accountability reconfiguration). Single-coder analysis supported by a dated coding manual and audit trail; disconfirming evidence was actively sought.

Limitations

• Stated: single national context limits contextual generalisability (analytical rather than empirical generalisation); the coordinated, compressed RCoC rollout plus Kenyan capacity constraints and regulatory gaps may have made default compliance and epistemic authority concentration especially visible; the regulatory-gap severity is likely context-dependent, while the persuasive authority of narrative outputs and the asymmetric burden of justification may travel more readily.
• Stated: reliance on interview data captures accounts rather than direct observation of decision-making; the study reflects officials' perspectives only, with no citizens interviewed, so trust claims are perceptions and inferences rather than citizen-level evidence.
• Evident in the text: analysis was conducted by a single researcher without independent coding; the study was not designed to compare generative with non-generative systems, so the claim that GenAI produces stronger authority effects than earlier technologies is not empirically established; the accountability-preserving counterexample rests on a single case.