Making AI reviewable and usable in government can backfire: codified procedures improve oversight on paper but create a playbook for successors to legally entrench politicized decisions, increasing strategic exploitation and locking in automation expansions.

AI Governance under Political Turnover: The Alignment Surface of Compliance Design

Andrew J. Peterson · April 22, 2026 · arXiv (Cornell University)

openalex theoretical n/a evidence 7/10 relevance Source PDF

A formal model shows that embedding probabilistic AI in codified, reviewable administrative procedures can improve detectability of unlawful departures yet also create stable approval boundaries that political successors learn to exploit, increasing vulnerability to strategic misuse and making rollbacks difficult.

Governments are increasingly interested in using AI to make administrative decisions cheaper, more scalable, and more consistent. But for probabilistic AI to be incorporated into public administration it must be embedded in a compliance layer that makes decisions reviewable, repeatable, and legally defensible. That layer can improve oversight by making departures from law easier to detect. But it can also create a stable approval boundary that political successors learn to navigate while preserving the appearance of lawful administration. We develop a formal model in which institutions choose the scale of automation, the degree of codification, and safeguards on iterative use. The model shows when these systems become vulnerable to strategic use from within government, why reforms that initially improve oversight can later increase that vulnerability, and why expansions in AI use may be difficult to unwind. Making AI usable can thus make procedures easier for future governments to learn and exploit.

Summary

Main Finding

Peterson shows that making AI-assisted administrative decision‑making more auditable and standardized can paradoxically increase the risk of procedurally clean democratic erosion under political turnover. By stabilizing a repeatable “pass-or-flag” review boundary (the paper’s “alignment surface”), compliance infrastructures become learnable and thus exploitable by successor actors. Reforms that improve oversight in the short run can, beyond a threshold, make within-form abuse easier and harder to unwind later.

Key Points

Alignment surface: the inherited, operational boundary created by rules, logs, thresholds, and review procedures that maps administrative actions into binary screening outcomes (Pass vs Flag). It is the practical, learnable interface through which compliance is judged.
Learnability mechanism: standardization and auditability make what clears review predictable; that predictability lowers the search/coordination cost for a successor seeking to pursue legally plausible but substantively abusive actions.
Model structure: three-stage game
- t = 0: slow-moving safeguards (statute, institutional design, investments) are set;
- t = 1: agencies choose AI adoption architecture (scale x and codification s) under modernization pressure;
- t = 2: a successor (democratic with prob. 1−δ, autocratic with prob. δ) inherits the system and may attempt overt or within-form undermining.
Design variables: scale x (breadth/depth of automation), codification s (degree of standardization/reusability across cases), safeguards r (throughput limits, separation of review/deployment, access controls).
Main formal results (informal):
- A threshold condition identifies when an inherited compliance architecture becomes politically exploitable.
- Codification dilemma (codification flip): once blatant violations are deterred, further increases in auditability plus internal standardization can increase vulnerability because they reduce the cost of locating legally defensible exploits.
- Pressure-and-persistence: modernization episodes that expand scale/codification can create durable, path‑dependent vulnerabilities that are costly or slow to unwind after the episode passes.
Distinction emphasized: auditability/transparency are necessary but not sufficient for democratic robustness—visibility to outsiders also implies legibility to insiders.
Relation to other concepts: connects to bureaucratic control/legibility literatures, autocratic legalism/constitutional hardball literature, and ML concepts like specification gaming and reward hacking.

Data & Methods

Method: formal theoretical model (game-theoretic, three-stage) that formalizes adoption choices and successor incentives under uncertainty about regime type (probability δ of autocratic successor).
Core formal elements:
- Action space A partitioned into P (permissible), U (ambiguous), I (impermissible).
- Operational compliance mapping Rs(a) ∈ {Pass, Flag} that defines the alignment surface.
- Agency choices: scale x and codification s; upstream safeguards r set at t = 0.
- Successor payoff tradeoffs incorporate detection/contestability costs under inherited enforcement institutions.
No empirical dataset is analyzed in the paper; the argument is theoretical and conceptual, with suggested empirical signatures and policy implications for future work and testing.
Extensions: a short post-crisis repair extension analyzes whether temporary modernization is later unwound; discussion offers design and empirical signatures.

Implications for AI Economics

Adoption decisions must internalize political externalities: economic efficiency gains from AI (lower labor costs, scalability) come with longer-run political risk from path-dependent compliance architectures. Cost–benefit analyses should include expected value of future regime types and the exploitability threshold.
Investment lock‑in and sunk-cost dynamics: large-scale modernizations (high x) create durable capital and workflows that are costly to reverse, implying strong path dependence and potential over‑investment in modes that increase future political fragility.
Trade-offs in procurement and regulation:
- Standard metrics (auditability, reproducibility) reduce oversight cost but raise insider learnability; regulators and procurement officers should balance short-run efficiency against long-run democratic robustness.
- Mitigation levers include limiting codification (reduce s), restricting operational scale (limit x), preserving heterogeneity across offices, restricting access rights, introducing external, independent review layers, and designing review artifacts to be less mechanically exploitable (e.g., randomized or multi-source attestations, non-deterministic checks).
Social planner vs incumbent incentives: incumbents facing modernization pressure may rationally choose architectures that improve present legitimacy/efficiency but raise the expected social cost under possible adversarial heirs; this creates a role for ex‑ante safeguards (r) and investment rules that internalize intertemporal political risk.
Measurement and empirical agenda for AI economics:
- Construct indices of automation scale (x), codification (s), and safeguard strength (r) to study correlation with subsequent within-form administrative abuse or legalistic consolidation.
- Study natural experiments where agencies adopted AI-driven workflows and later experienced political turnover to test the pressure‑and‑persistence prediction.
- Model calibration: incorporate probabilities of autocratic succession (δ) into cost‑benefit frameworks for public AI investments.
Policy recommendation summary: incorporate turnover-aware risk into cost–benefit and procurement frameworks, strengthen upstream safeguards that reduce insider learnability (not just external log visibility), retain institutional heterogeneity, and mandate independent, externalized audit/reversal capacity to reduce the payoff to within-form exploitation.

Assessment

Paper Typetheoretical Evidence Strengthn/a — The contribution is purely theoretical: it generates logical implications from a formal model but presents no empirical tests, randomized variation, or quasi-experimental identification to establish causal effects in data. Methods Rigormedium — The work appears to use a formal model to produce internally consistent comparative statics and mechanisms, which supports strong internal validity; however, the rating is conservative because rigor depends on the clarity of assumptions, robustness checks, and potential alternative specifications, and the paper lacks empirical validation or calibration. SampleNo empirical sample — the paper analyzes an abstract model with stylized agents (institutions, political actors, automated decision systems) choosing scale of automation, degree of codification, and safeguards; outcomes are derived analytically rather than estimated from data. Themesgovernance adoption org_design human_ai_collab IdentificationNone — the paper uses a formal, game-theoretic/analytical model to derive comparative static predictions; causal claims follow from model structure and assumptions rather than empirical identification. GeneralizabilityAbstract model may not capture real-world legal and institutional heterogeneity across jurisdictions, Relies on stylized assumptions (e.g., rational actors, specific information structures) that may not hold in practice, No empirical calibration — quantitative magnitudes and thresholds are not tied to observed contexts, Ignores operational frictions and enforcement costs that could alter strategic incentives, Limited attention to behavioral or bounded-rational responses by bureaucrats and courts

Claims (9)

Claim	Direction	Confidence	Outcome	Details
Governments are increasingly interested in using AI to make administrative decisions cheaper, more scalable, and more consistent. Adoption Rate	positive	high	government interest in AI adoption for administrative decisions (cost, scale, consistency)	0.06
For probabilistic AI to be incorporated into public administration it must be embedded in a compliance layer that makes decisions reviewable, repeatable, and legally defensible. Regulatory Compliance	positive	high	requirements for legal/administrative incorporation of probabilistic AI	0.02
That compliance layer can improve oversight by making departures from law easier to detect. Governance And Regulation	positive	high	detectability of departures from law (oversight effectiveness)	0.02
The compliance layer can also create a stable approval boundary that political successors learn to navigate while preserving the appearance of lawful administration. Governance And Regulation	negative	high	creation of a stable approval boundary exploitable by successive governments	0.02
We develop a formal model in which institutions choose the scale of automation, the degree of codification, and safeguards on iterative use. Governance And Regulation	null_result	high	institutional choices regarding automation scale, codification, and safeguards (model inputs/decision variables)	0.02
The model shows when these systems become vulnerable to strategic use from within government. Governance And Regulation	negative	high	vulnerability of automated systems to strategic internal use	0.02
The model explains why reforms that initially improve oversight can later increase that vulnerability. Governance And Regulation	negative	high	long-run effect of oversight-improving reforms on system vulnerability	0.02
The model shows why expansions in AI use may be difficult to unwind. Adoption Rate	negative	high	persistence/irreversibility of AI adoption (difficulty of unwinding expansions)	0.02
Making AI usable can thus make procedures easier for future governments to learn and exploit. Governance And Regulation	negative	high	ease with which future governments can learn and exploit administrative procedures enabled by AI	0.02