# Webinar: Regulating at the Speed of AI with Francis Fukuyama and Nate Persily

## Executive summary

This webinar discusses the urgent need for comprehensive AI governance, moving beyond simple application-level regulation to address systemic, technological risks posed by frontier models. Speakers argue that AI regulation is inevitable, requiring a blend of adapting existing legal frameworks (e.g., tort law) and establishing new, specialized regulatory bodies. Key challenges include the global nature of the technology (US vs. China), the difficulty of enforcing rules across international borders, and the need for technical capacity within government agencies to manage pre-deployment testing and monitor advanced models.

## Key takeaways

- AI Governance Must Regulate Technology, Not Just Applications: While existing regulations cover AI use by banks or pharma, the discussion must shift to regulating the technology itself, particularly the major frontier models, due to concerns about existential risk and loss of control. The goal is to chart a safe AI future by adapting the existing legal order to deal with AI's systemic impact.
- Self-Regulation and Existing Law Are Insufficient: Relying on AI companies to 'pace the frontier' is risky, as evidenced by incidents like the Hugging Face event. While existing tort law could punish misuse, the potential costs associated with new risks (e.g., SEAB risks, loss of control, cyber risks) are so huge that current litigation liability regimes may not be sufficient.
- A Specialized, Multi-Layered Regulatory Approach is Needed: Due to AI's fast-moving and complex nature, the speakers suggest that a specialized regulator is necessary, rather than folding AI into existing agencies. This body would need technical capacity to conduct pre-deployment testing and monitor models at the 'model layer,' while enforcement requires a 'whole-of-government approach' across all sectors (e.g., energy, finance).
- The Open Model Challenge and Geopolitics: The rise of powerful open models (especially from China) poses a unique regulatory problem. Since these models can be adapted globally and are hosted in the cloud, it becomes increasingly difficult to attribute harm to a single developer, complicating regulatory oversight compared to proprietary, closed models.

## Technical details

- Model Safety and Testing: The need for a regime of pre-deployment testing, requiring AI labs to obtain federal stamp of approval. This requires a new agency with technical capacity to evaluate models at the 'model layer.' The concept of 'sandboxes' (e.g., for research) must be locked down to prevent internet access and unauthorized deployment.
- AI Infrastructure and Compute: The massive investment in data centers is propping up the global economy. The discussion highlights the importance of compute infrastructure and the geopolitical competition for AI dominance, particularly concerning semiconductor supply chains (e.g., TSMC).
- Open vs. Closed Models: Open models (like those on Hugging Face) offer adaptability and lower cost, making them popular for middle powers and developing nations. However, they complicate regulatory oversight because the source of harm is diffuse, unlike proprietary, closed models.
- Regulatory Precedents: Historical examples of successful regulation include the FAA and the creation of specialized bodies (like the Interstate Commerce Commission, which was deemed unworkable for modern tech). The EU's AI Act is noted as the most comprehensive regulation to date.

## Practical implications

- For build engineers, the focus must shift to designing systems with inherent safety controls and robust sandboxing mechanisms to mitigate 'loss of control' risks from advanced models.
- Infrastructure planning must account for the geopolitical competition in compute power, recognizing that reliable access to advanced chips and energy is critical for deployment.
- Compliance strategies must anticipate a shift toward mandatory pre-deployment testing and model-level certification, requiring detailed documentation of model provenance and training data.
- The need for 'AI management' roles suggests that technical teams will require specialized skills not just in building, but in overseeing and managing autonomous AI agents.

## Topics

AI Governance, Geopolitics, Systemic Risk, Compute Infrastructure, Regulatory Law, AI Power and New Rules of Global Strategy

Source: https://www.youtube.com/watch?v=dAzTioGQIzQ