US-China AI compute disclosure: a quasi-linear mechanism with mixed audits

Abstract

The growth and development of Artificial Intelligence (AI) in the past decades has generated both optimism and fear, as while these technologies possess great potential for good, they could also be used to further national interests, including through war. One limiting factor in the international regulation of AI is that often states and companies do not disclose the full extent of their progress. We utilize the mechanism design approach to study whether a US-China “compute disclosure” treaty can be made incentive compatible and individually rational when states have private information, valuable outside options to race, and audits are noisy. We propose a directrevelation mechanism with a per-unit levy on reported compute 𝜏, a linear deviation penalty 𝜅, a mixed audit of strength 𝑞, and report-independent transfers 𝑠. Theoretically, we show that adding a point mass 𝑞 > 0 of perfect audits to otherwise continuous noise restores first-order deterrence, and truth-telling holds whenever |𝐵 ′𝑖 (𝑥𝑖) − 𝛽𝑖𝐾 𝑝′ fail(𝑦) − 𝜏| ≤ 𝜅𝑞. We derive closedform individual rationality conditions, minimal transfers 𝑠 min 𝑖 , and an “audit-noise frontier” characterizing feasibility. Monte-Carlo simulations are used to compare the exogenous- and endogenous-compute models and their sensitivities to key parameters. In the endogenous case, we find that re-centering 𝜏 and strengthening 𝜅𝑞 lowers total compute 𝑦, reduces existential risk exposure, and increases welfare. We find that implementing the mechanism lowers E[𝑝fail(𝑦)], the risk of existential collapse caused by AI, by 26% and increases welfa re, and that the mechanism implements truthful reporting and attains individual rationality 90% of the time with a subsidy of 𝑠 ∗ ≈ 3.736, worth approximately $10 billion. Finally, we discuss the robustness of the solutions found, the impact of institutional constraints and outline potential extensions, such as games with repeated interaction and/or multi-dimensional types.Η διατριβή “Γνωστοποίηση υπολογιστικής ισχύος ΤΝ ΗΠΑ–Κίνας: Ένας μηχανισμός ψευδογραμμικής χρησιμότητας με μεικτούς ελέγχους” υποβλήθηκε προς εκπλήρωση των απαιτήσεων για την απόκτηση μεταπτυχιακού τίτλου σπουδών στα Οικονομικά.