Import AI 456: RSI and economic growth; radical optionality for AI regulation; and a neural computer

Researchers at the Institute for Law & AI propose the concept of 'Radical Optionality' as a third path for AI regulation. The core idea is to avoid overregulation in the short term while rapidly building institutions, information channels, and legal authorities so that democratic governments can still make good decisions about transformative AI systems as the environment evolves.

Given the enormous risks of AI development, the authors argue that 'governments should be willing to spend vast sums of money, energy, and political capital to preserve optionality.' They emphasize that governments should be wary of counterproductive interventions but need not worry too much about the actual monetary cost of any realistic measure that seems likely to have a net positive effect.

Information-gathering authorities: including transparency requirements (companies publicly disclose AI system information), reporting requirements (share information with government agencies), and establishing auditing systems to verify information accuracy.

Whistleblower protections: ensure that frontier lab employees can report risk information.

Information sharing within and outside government: enhance intergovernmental coordination, especially regarding sensitive information about AI progress and critical supply chain protection.

Flexible rules and definitions: avoid premature regulation, possibly using conditional 'if-then' regulatory commitments or setting high-level goals (e.g., risk reduction) that companies define how to achieve.

Assessment and evaluation: develop government and third-party capabilities to assess AI capabilities and safety.

Model weights and algorithmic secret security: strengthen protection of neural network weights and algorithmic secrets by promoting voluntary physical and cybersecurity standards.

Talent recruitment: increase investment in technical talent for institutions like UK AISI and US CAISI, which underpins all other interventions.

To the criticism of 'encouraging aggressive regulatory action,' the authors argue that the above recommendations 'are not substantive authorities that can be easily abused.' However, Import AI's editor notes that a sufficiently motivated government could implement authorities more forcefully than the drafters envisioned.

On democratic legitimacy: flexibility may require weakening some democratic procedures, such as allowing agencies to waive notice-and-comment periods for certain rulemakings.

On power concentration and government abuse: the authors are 'basically convinced' that the risk of government control over AI development is significant, so they do not recommend large-scale expansion of emergency powers like the Defense Production Act. One mitigation measure is to require the government to 'only use law-abiding AI systems.'

On private governance: although the authors support the idea of a 'regulatory market,' they believe any governance that primarily relies on the private sector (e.g., independent verification bodies) ultimately requires the government to have basic technical capabilities.

Import AI editor Jack Clark says he agrees with all the recommendations and notes that he has advocated for many of these measures himself in recent years. He believes society can do more to prepare for the massive changes ahead.

The authors conclude: 'The costs of implementing these policies are modest relative to the potential benefits. Conversely, the cost of inaction could be catastrophic.' The editor agrees.