[Paper Review] Computing Power and the Governance of Artificial Intelligence

Girish Sastry, Lennart Heim+17 more|arXiv (Cornell University)|Feb 13, 2024

Ethics and Social Impacts of AI20 citations

TL;DR

The paper argues that governing AI compute can meaningfully enhance AI governance by improving regulatory visibility, resource allocation toward safe uses, and enforcement against misuse, while outlining guardrails and risks.

ABSTRACT

Computing power, or "compute," is crucial for the development and deployment of artificial intelligence (AI) capabilities. As a result, governments and companies have started to leverage compute as a means to govern AI. For example, governments are investing in domestic compute capacity, controlling the flow of compute to competing countries, and subsidizing compute access to certain sectors. However, these efforts only scratch the surface of how compute can be used to govern AI development and deployment. Relative to other key inputs to AI (data and algorithms), AI-relevant compute is a particularly effective point of intervention: it is detectable, excludable, and quantifiable, and is produced via an extremely concentrated supply chain. These characteristics, alongside the singular importance of compute for cutting-edge AI models, suggest that governing compute can contribute to achieving common policy objectives, such as ensuring the safety and beneficial use of AI. More precisely, policymakers could use compute to facilitate regulatory visibility of AI, allocate resources to promote beneficial outcomes, and enforce restrictions against irresponsible or malicious AI development and usage. However, while compute-based policies and technologies have the potential to assist in these areas, there is significant variation in their readiness for implementation. Some ideas are currently being piloted, while others are hindered by the need for fundamental research. Furthermore, naive or poorly scoped approaches to compute governance carry significant risks in areas like privacy, economic impacts, and centralization of power. We end by suggesting guardrails to minimize these risks from compute governance.

Motivation & Objective

• Motivate the study of compute as a lever for AI governance and articulate why compute is a particularly effective intervention.
• Define the AI triad (data, algorithms, compute) and explain compute’s role in frontier models and deployment.
• Outline how compute governance can enhance three governance capacities: visibility, allocation, and enforcement.
• Identify potential risks and guardrails for compute governance and discuss current and future policy options.

Proposed method

• Develop a conceptual framework framing compute as a governance lever based on its detectability, excludability, quantifiability, and supply-chain concentration.
• Characterize the AI lifecycle (design, training, enhancement, deployment) and map compute footprints to each stage.
• Review existing compute governance efforts (domestic compute capacity, export controls, reporting thresholds) and national/international policy instruments.
• Propose illustrative governance mechanisms across visibility, allocation, and enforcement categories with design considerations.
• Discuss risks of compute governance (privacy, centralization, feasibility) and propose guardrails to mitigate them.

Figure 1 : Summary of the core concepts in the report. Compute is attractive for policymaking because of four properties. These properties can be leveraged to design and implement policies that enable three critical capacities for the governance of AI.

Experimental results

Research questions

• RQ1 How can compute governance be used to increase regulatory visibility into AI capabilities and use?
• RQ2 What policy mechanisms can reallocate or manage compute to promote safe and beneficial AI development?
• RQ3 How can compute be used to enforce norms and regulations on AI development and deployment?
• RQ4 What guardrails are necessary to minimize privacy and centralization risks in compute governance?
• RQ5 What are the practical challenges and readiness levels of compute-based governance approaches?

Key findings

• Compute is highly impactful for frontier AI models due to its centralized, tangible nature and measurable footprint.
• Compute governance can enhance three capacities of AI governance: visibility, allocation, and enforcement.
• A variety of illustrative mechanisms exist for visibility (e.g., reporting requirements, AI chip registries), allocation (e.g., targeted support for beneficial AI, pace adjustments), and enforcement (e.g., hardware-based controls, norm enforcement).
• There are significant risks and feasibility concerns, including privacy, data leakage, and potential centralization of power, necessitating guardrails such as excluding small-scale compute and emphasizing privacy-preserving practices.
• Current governance efforts (domestic capacity, export controls, reporting thresholds) illustrate both momentum and tensions, highlighting the need for a holistic theory and careful design of compute-based policies.

Figure 2 : The AI Triad. The three key technical inputs to AI are data, algorithms, and compute. Human capital is required for all inputs.