AI Governance Needs Radical Optionality

Abstract

Governments should avoid over-regulation in the short term while building the institutional capacity needed to competently regulate transformative AI systems if and when they arrive. The point is to maximise optionality — equipping institutions with tools that can respond to a wide range of foreseen and unforeseen developments — which shows that security and innovation are compatible rather than conflicting priorities.

Context

• Guest commentary in AI Frontiers by Charlie Bullock, Senior Research Fellow at the Institute for Law & AI, summarising a longer essay co-authored with Christoph Winter (published at radical-optionality.ai).
• Framed against two existing poles:
  • Permissionless innovation — a libertarian position limiting government to enforcing existing law and facilitating industry self-regulation through “soft law” tools such as voluntary standard-setting.
  • The precautionary principle — heavy-handed regulation restricting frontier AI development until developers can prove their systems are safe.
• Radical optionality is offered as a third path between them.

The argument under uncertainty

• The argument is addressed to anyone who thinks there is even a small chance that predictions of AGI, superintelligence, or “powerful AI” materialise; the author declines to argue with those certain it is all hype.
• AI’s future impacts are highly uncertain. Assuming some possibility of transformative systems within roughly 15 years, we remain uncertain about timing, mechanism, characteristics, benefits, and risks. Plausible futures span:
  • systems that are mostly harmless and highly beneficial, because commercial incentives push toward safety and alignment;
  • systems that are dangerous and hard to control, possibly extinction-capable absent guardrails;
  • securitisation, with the U.S. government developing systems behind closed doors as a clandestine military project;
  • decentralised, democratic development.
• The apparent tradeoff. Restrictive regulation slows innovation and forgoes benefits; well-designed regulation might mitigate real risks. Only ideologues claim certainty about a poorly understood emerging technology, so decisions must be made under substantial uncertainty.
• The reframing. That tradeoff framing misses that some measures increase security at no significant cost to innovation — these are the measures optionality-maximising governance prioritises.

What optionality-increasing measures look like

Information-gathering authorities

• Light-touch tools at the top of the list: whistleblower protections, reporting requirements, and transparency mandates.
• Rationale: “information is the lifeblood of good governance.” Access to information about AI risks — and agency experience in securely processing and interpreting it — is a foundational building block for later governance.
• Equally foundational: mechanisms for sharing information securely and intelligently within government, and where appropriate between governments.

Direct capacity building

• Above all, enabling regulatory bodies to hire and retain elite talent.
• Benchmark for what taking transformative AI seriously looks like: Meta’s hiring spree with ten-figure compensation offers for top AI researchers.
• Governments cannot match private-sector salaries, but hiring and contracting reforms are needed in both the U.S. and the EU.
• Cited proof of concept: the UK AI Security Institute, which receives ten times the funding of its U.S. counterpart despite the UK’s smaller GDP and industry relevance.

Other measures (from the full essay)

• Incentivising lab security.
• Avoiding premature and overbroad federal preemption of state AI laws.
• Building out an ecosystem for model assessments and evaluations.

Compatibility with other governance proposals

• Private governance (Dean Ball; Gillian Hadfield and Jack Clark): government certifies an ecosystem of competing private regulators offering nimble, opt-in regulatory services.
• Tort liability (Gabriel Weil): insurance requirements, punitive damages, and strict liability for certain harms, forcing developers to internalise the risks their products generate.
• Management-based regulation (Cary Coglianese): requiring risk mitigation while leaving companies broad discretion over which measures to adopt.
• The author treats all three as consistent with and useful for maintaining and increasing optionality.

Anticipated objections

• From the precautionary side: if one is confident that restrictive regulation’s safety benefits outweigh the innovation costs, radical optionality does not go far enough — though it would still beat the status quo.
• From the libertarian side: building regulatory capacity while promising restraint may look like “giving the government a hammer and promising that agencies will not start hallucinating nails.”
• The author concedes he cannot promise new authorities will never be abused, nor that people will agree on when dual-use systems become advanced enough that regulation is a national security imperative.
• Counter-consideration: if capability progress triggers a surge in public demand for regulation, companies may themselves prefer a government able to regulate competently and in a targeted way rather than reactively.

Closing position

• Radical optionality functions as the organising principle for the Institute for Law & AI’s research and consulting: which projects to take on, which bills to comment on, and which policies to push are judged substantially by what maximises optionality.
• The stated aim of the piece is to persuade readers to adopt the framing and to stop treating security and innovation as conflicting priorities.