Back
Executive Order 14110
webCredibility Rating
4/5
High(4)High quality. Established institution or organization with editorial oversight and accountability.
Rating inherited from publication venue: RAND Corporation
Published June 2025 by RAND's Technology and Security Policy Center, this report is directly relevant to implementing compute governance provisions similar to those in Executive Order 14110, and informs debates about the feasibility of compute-based AI monitoring regimes.
Metadata
Importance: 62/100organizational reportanalysis
Summary
This RAND Corporation report uses game-theoretic analysis to examine detection and monitoring mechanisms cloud service providers could use to identify large AI training runs. It finds significant gaps in monitoring schemes based solely on floating-point operation thresholds, showing many capable AI models could be trained without detection. The authors recommend hybrid compute and non-compute governance approaches and continued research into adaptive monitoring frameworks.
Key Points
- •FLOPs-only monitoring thresholds create significant detection gaps; many capable AI models could be trained without triggering reporting requirements.
- •Game-theoretic modeling reveals strategic vulnerabilities in compute governance schemes that adversarial actors could exploit.
- •Cloud service providers need more sophisticated detection strategies beyond simple compute thresholds to effectively monitor AI training runs.
- •Policymakers should pursue hybrid approaches combining compute-based and non-compute-based AI governance mechanisms.
- •Ongoing research into detection vulnerabilities and adaptive thresholds is necessary as AI capabilities and training methods evolve.
Cited by 1 page
| Page | Type | Quality |
|---|---|---|
| Compute Monitoring | Approach | 69.0 |
Cached Content Preview
HTTP 200Fetched Mar 15, 20265 KB
Strategies and Detection Gaps in a Game-Theoretic Model of Compute Governance | RAND
Skip to page content
This report documents research and analysis conducted as part of a study to investigate detection and monitoring mechanisms that cloud service providers could employ to identify large artificial intelligence (AI) training runs. The intended audience of this report is national policymakers interested in compute-based AI governance. This report may be of interest to cloud service providers and infrastructure-as-a-service companies.
Strategies and Detection Gaps in a Game-Theoretic Model of Compute Governance
Alvin Moon , Padmaja Vedula , Jesse Geneson , Simon Bar-on
Research Published Jun 16, 2025
Download PDF
Share on LinkedIn
Share on X
Share on Facebook
Email
This report documents research and analysis conducted as part of a study to investigate detection and monitoring mechanisms that cloud service providers could employ to identify large artificial intelligence (AI) training runs. The intended audience of this report is national policymakers interested in compute-based AI governance. This report may be of interest to cloud service providers and infrastructure-as-a-service companies.
Key Findings
Cloud service providers might not be able to report and detect AI training if they are only obligated to monitor and report activity based on floating-point operation thresholds in future AI governance policy.
In the future, many types of capable AI models will likely exist whose training will be hard to detect. The authors outline strategies for cloud service providers that could aid with AI training detection in cloud computing environments.
Recommendations
To develop effective AI governance, policymakers should support efforts to find detection gaps in compute-based monitoring schemes.
Policymakers should continue to pursue both compute- and noncompute-based AI governance.
Continuing research into effective thresholds for compute monitoring is required to create a robust compute-based monitoring framework that can adapt to technological progress.
Subscribe to the Policy Currents newsletter
Email
Subscribe
Topics
Artificial Intelligence
Network Analysis
Document Details
Copyright: RAND Corporation
Availability: Web-Only
Year: 2025
Pages: 25
DOI: https://doi.org/10.7249/RRA3686-1
Document Number: RR-A3686-1
Citation
RAND Style Manual
Moon, Alvin, Padmaja Vedula, Jesse Geneson, and Simon Bar-on, Strategies and Detection Gaps in a Game-Theoretic Model of Compute Governance, RAND Corporation, RR-A3686-1, 2025. As of February 25, 2026: https://www.rand.org/pubs/research_reports/RRA368
... (truncated, 5 KB total)Resource ID:
1e392ce476b43c8f | Stable ID: ODYxYzU5Yj