Working Papers

Title: The administrative burden of responsible AI and governed autonomy in research organisations

Author: Steve J. Bickley, Benno Torgler

Abstract: Responsible artificial intelligence (AI) requires organisational oversight, but access procedures can delay legitimate work, transfer costs to employees, discourage disclosure and widen inequalities. We integrate administrative-burden theory, enabling bureaucracy, shadow innovation and sociotechnical accountability to examine this problem in research organisations. A Governance Friction Model links safeguards, procedural costs, financial support and disclosure concerns to sanctioned use, shadow use and non-use. Hypothetical simulations show that total loss can rise even while direct AI-related harm falls. Reducing procedural burden while preserving safeguards can lower loss and support greater governance intensity, whereas funding a burdensome pathway can increase both sanctioned participation and total loss. These conditional results motivate Governed AI Access: an institutional architecture that governs the conditions of experimentation through reusable permissions and review when material risks change. An AI experimentation passport connects training and bounded autonomy to eligible providers, institutional payment, proportionate disclosure and escalation. A universal access floor, shared capacity and responsive top-ups are designed to accommodate heterogeneous demand through justified allocations beyond a common floor. The architecture retains continuing provider assurance, contestable decisions, protections for affected people and the freedom to decline unsuitable AI use. Seven propositions identify conditions under which it could succeed or fail. The contribution is to make access design and its behavioural consequences part of responsible AI governance.