As deletion requests climb, auditors can safely cut inspection intensity because degraded unlearning makes non-compliance easier to detect; and while secret audits reveal more, they can paradoxically reduce regulatory cost‑effectiveness compared with public auditing.

Main Finding

The paper develops the first economic framework for auditing machine unlearning (MU) compliance by combining (ε, δ)-certified unlearning theory with regulatory enforcement. It (i) quantifies the auditor’s probabilistic detection capability via a hypothesis-testing interpretation of certified unlearning, (ii) models the strategic interaction between an AI operator and an auditor with MU-specific utility–compliance tradeoffs, and (iii) shows—both analytically and empirically—several counterintuitive regulatory insights: (a) auditors may optimally reduce inspection intensity as deletion requests increase because weaker unlearning makes non-compliance easier to detect, and (b) disclosed (transparent) auditing yields superior regulatory cost-effectiveness and produces win–win outcomes versus undisclosed auditing, despite undisclosed audits offering informational advantages.

Key Points

• Novel integration: First economic auditing framework tailored to MU that embeds (ε, δ)-certified unlearning into a game-theoretic regulatory model.
• Operator decision: The AI operator chooses an unlearning certification level ε ≥ 0 (smaller ε = stricter unlearning) to trade off model utility against audit risk.
• Utility model: Model utility (negative test loss) is parameterized as u(ε) = −G · η^2 / ε, where η is the fraction of deleted samples and G aggregates model constants (dimension, Lipschitz/convexity constants, privacy failure δ).
• Auditor verification: Each inspection is modeled as an independent hypothesis test comparing (i) a model trained without deleted data (null) and (ii) a model trained on full data then unlearned (alternative). (ε, δ)-certified unlearning implies probabilistic indistinguishability and sets detection limits.
• Audit process: Two-stage audit—investigation (m independent tests, zero-tolerance: any failure = non-compliance) and enforcement (fine + corrective unlearning).
• Game structure: Simultaneous game for undisclosed auditing; sequential (auditor commits to inspection intensity publicly) for disclosed auditing.
• Analytical challenge & solution: MU-specific nonlinearities in utility and detection probabilities create a bivariate nonlinear fixed-point. The authors develop an auxiliary transformation that reduces it to a univariate problem, enabling proofs of equilibrium existence, uniqueness, and structural properties without closed-form solutions.
• Regulatory paradoxes:
  • Unlearning audit paradox: As the volume of deletion requests rises, optimal inspection intensity can decrease because the operator’s incentive to weaken unlearning (to preserve utility) makes non-compliance more detectable.
  • Auditing transparency paradox: Although undisclosed auditing gives auditing informational advantages in inspection-time, disclosed auditing (via commitment) is more cost-effective for enforcement and improves payoffs for both auditor and operator.
• Empirical findings: Experiments with real MU implementations show disclosed auditing increases the auditor’s payoff by up to 2549.30% and the operator’s payoff by up to 74.60% relative to a state-of-the-art benchmark; disclosed auditing also outperforms undisclosed auditing for both players.

Data & Methods

• Theoretical foundations:
  • Uses (ε, δ)-certified unlearning literature to relate ε to indistinguishability between unlearned and retrained models.
  • Interprets certified unlearning through hypothesis-testing to derive detection probabilities for auditor inspections.
• Modeling:
  • Operator utility specified analytically as u(ε) = −G η^2 / ε, capturing how stricter unlearning (small ε) increases test loss.
  • Auditor conducts m independent probabilistic tests in an investigation stage, applying a zero-tolerance rule (any failed test triggers enforcement).
  • Payoffs include enforcement fines and corrective unlearning costs.
  • Two game forms: simultaneous (undisclosed auditing) and sequential with commitment (disclosed auditing).
• Analytical technique:
  • Introduces an auxiliary transformation to convert a bivariate nonlinear fixed-point problem into a univariate formulation; used to prove existence, uniqueness, and monotonicity/structure of equilibrium without closed-form solutions.
• Experiments:
  • Real-data MU experiments validate the inverse relation between ε and model performance predicted by theory.
  • Numerical simulations compare disclosed vs undisclosed auditing and a benchmark, measuring auditor and operator payoffs under varying deletion request volumes, inspection costs, fines, and η.

Implications for AI Economics

• Enforcement design must account for probabilistic verification: MU’s statistical nature fundamentally changes detection dynamics compared to deterministic deletion auditing. Economists and policymakers should model detection as probabilistic and incorporate it into penalty and inspection design.
• Inspection intensity is not monotone in deletion volume: Regulators facing growing deletion requests might optimally reduce per-request inspection intensity because larger-scale deletion demands incentivize operators to weaken unlearning (which makes non-compliance easier to detect). This provides an economic explanation for observed reductions in inspection intensity despite rising deletion volumes.
• Transparency as a commitment device: Making audit regimes and inspection intensities public (disclosed auditing) can improve cost-effectiveness and generate higher payoffs for both auditors and operators relative to undisclosed audits—so commitment/transparency mechanisms can be welfare-improving in AI governance.
• Tradeoffs for operators: Operators will strategically choose ε considering both accuracy loss and audit risk. Policy instruments (fine levels, corrective unlearning mandates) can be tuned to shift operators toward stricter unlearning without excessive inspections.
• Practical policy guidance: Regulatory frameworks (e.g., GDPR, PIPL, AI Act) should incorporate certified unlearning concepts when crafting compliance-verification rules and consider allowing or mandating disclosure/commitment of inspection regimes to increase enforcement effectiveness.
• Research directions: Future economic models should relax some assumptions (zero-tolerance, independence of tests, exact form of u(ε)), incorporate multi-operator/multi-auditor settings, costs of public disclosure, and empirical calibration across model families and MU algorithms.

Limitations to note (for economic interpretation): the utility function and detection mapping derive from specific certified-unlearning bounds and model assumptions; real-world MU implementations and audit inspection procedures may differ in dependence structure and cost; the zero-tolerance rule is a modeling choice that could be softened in alternative regulatory designs.