Page 34 - AI Governance Day - From Principles to Implementation

AI Governance Day - From Principles to Implementation

"Compute governance", the setting of rules on computing resources to achieve governance, can
be an attractive tool for AI governance. This is because compute is detectable and quantifiable,
allowing for effective monitoring and control. For example, energy-intensive, specialized data
center infrastructure is an indicator of compute activity. In contrast, while data and algorithms
are also essential ingredients of AI, it is much more challenging for governments to quantify
them.

Many are of the opinion that using compute providers (e.g. Microsoft Azure, Amazon Web
Services (AWS), Apple, Bytedance, Meta, Oracle, Tencent, and Google Cloud) as intermediary
regulators would be most effective in addressing risks associated with large-scale AI training to
prevent bad actors from training advanced AI models, rather than addressing all AI-related risks.
This is because non-compute-intensive AI models are often feasible to train and run on widely
available customer hardware, so cloud providers have less ability to oversee such activities.

Compute providers can therefore play an essential role in AI governance via four key functions:

• Securers: protecting AI systems and critical infrastructure
• Record keepers: improving transparency for regulators
• Verifiers: monitoring customer activities
• Enforcers: taking actions against breaches of rules

International cooperation is essential to handle cross-border supervision and data challenges
(e.g. ensuring that personal data is protected according to different regional standards
and regulations), as it reduces the risk of compute providers and AI developers moving to
jurisdictions with less regulatory oversight.

In addition to its potential role in regulation, compute has the potential to advance international
cooperation on AI, by enabling states and companies to demonstrate their adherence to their
commitments without leaking sensitive data. States may be able to show that approximately
all of their AI compute was used consistently with their commitments, meaning significant
compute would not have been available for other purposes. These approaches could leverage
Privacy-Enhancing Technologies (PETs) to enable assurance while preserving confidential data.

Potential discussion questions

• How does compute governance differ from data governance and algorithm governance?
• Are there real-world examples of effective compute governance?
• How can compute resources be effectively monitored and controlled to ensure
compliance with governance policies?
• How can compute providers improve transparency for regulators and stakeholders?
• What are the potential frameworks for international cooperation on compute governance?
• What are the potential risks of over-regulation, and how can they be mitigated?
• How might compute governance evolve with advancements in AI and computing
technologies?

4.4.3 Insights from the breakout sessions: theme 2

Implementing an AI governance framework involves addressing challenges across data,
compute, models, and deployment. Here is a direct and concise approach summarizing the
key points and elements discussed to implement an adequate framework.

24