AI Governance Reveals Compliance Design Vulnerabilities Under Political Turnover

A formal paper by Andrew J. Peterson, posted to arXiv on 22 April 2026, analyzes how embedding probabilistic AI into a procedural compliance layer for public administration changes incentives under political turnover. The model shows that design choices intended to improve reviewability and legal defensibility can create a stable approval boundary that successors learn to exploit, making systems vulnerable to strategic use and difficult to reverse.

The paper formulates an institutional choice problem where administrators set three design levers:

• •scale of automation, how much decision volume is shifted to AI;
• •degree of codification, how strictly rules and decision criteria are formalized;
• •safeguards on iterative use, controls on repeated or feedback-driven deployments.

The model treats the compliance layer as an observable approval boundary: higher codification increases detectability of deviations but may constrain discretionary reform. Iterative use with weak safeguards creates learning externalities for future actors, who can empirically discover approval paths that preserve apparent legality while achieving new political goals.

This work connects alignment thinking to public-administration literature and political-economy models of institutional drift. It reframes the alignment challenge from pure model behavior to the design of socio-technical compliance surfaces that interact with agent incentives. For ML practitioners and policy designers, the core insight is that usability and testability of AI systems are double-edged: making systems easier to audit and operate can lower the experimental cost for actors who seek to repurpose them.

The paper highlights a paradox: reforms that initially strengthen oversight by increasing codification or monitoring can later increase vulnerability because they create predictable approval criteria that opponents can exploit. This produces path dependence - expansions of automation become institutionalized and are costly to unwind. "Making AI usable can thus make procedures easier for future governments to learn and exploit," said Andrew J. Peterson.

Empirical follow-ups that map approval-boundary learning in deployed government pilots, and design research on safeguards that preserve auditability without creating predictable shortcuts, will be critical next steps.