Radical Optionality: Governing Transformative AI Under Uncertainty is a 2026 essay by Christoph Winter (Assistant Professor of Law and AI, University of Cambridge) and Charlie Bullock, both affiliated with the Institute for Law & AI. It argues for what the authors call "radical optionality": a governance strategy for transformative AI under which governments avoid overregulation in the short term while aggressively building the institutions, information channels, legal authorities, and expertise needed to respond competently to a range of future scenarios. It is a programmatic governance position rather than empirical evidence.
The essay was published by the Institute for Law & AI in 2026. The exact date is not given in the converted source file; it is dated approximately by internal references to the Claude Mythos Preview cyber capabilities and the June 2025 state-AI moratorium.
Summary of the argument
The authors frame the central problem of governing transformative AI as preserving democratic governments' ability to make good decisions as circumstances evolve, rather than committing now to a substantive regulatory regime. In the short term, on this account, that means avoiding overregulation while investing in the capacity to respond.
The argument rests on four stated assumptions: a real possibility of transformative AI (a transition "comparable to or more significant than the agricultural or industrial revolution") within roughly 10 years; profound uncertainty about its capabilities, risks, and the best governance responses; the proposition that a transformative dual-use national-security technology will inevitably require some government oversight; and the claim that building institutional capacity takes years, so society cannot wait until transformative capabilities have arrived.
The "radical" element is the authors' contention that governments should spend "an extraordinary amount of money, effort, and political capital" on optionality. They argue that even a 5% chance of positive impact on "the most important invention in human history" would justify the cost "a thousand times over in expectation," an expected-value argument rather than an empirical claim.
Positioning against alternatives
The essay positions radical optionality between and against four alternative approaches, which it treats as the central distinction it draws:
- "Let the market handle it" (Adam Thierer's "permissionless innovation"; Draghi's EU competitiveness review), rejected on the ground that a transformative dual-use technology with significant military applications will inevitably require a nonzero amount of regulation, and that "let the market handle it until national-security issues manifest" leaves government unprepared.
- The precautionary principle, of which the "hard" version (prohibit unless proven safe) is described as "simply bad policy," while a cost-benefit version is "not necessarily inconsistent with radical optionality."
- Anticipatory governance, rejected on the ground that governments are historically poor at predicting technological trajectories, citing the Audio Home Recording Act of 1992 against the hands-off early-Internet approach.
- "Muddling through" (Charles Lindblom's incrementalism), which the authors say shares their skepticism of grand regimes but preserves flexibility only by default, through inaction; radical optionality instead requires proactive investment in capacity.
Concrete policy suggestions
The essay sets out seven categories of capacity-building policy:
- Information-gathering authorities: transparency requirements (public disclosure; SB 53, NY RAISE Act) and reporting requirements (confidential agency reporting; the abandoned Biden-era BIS rule, EU GPAI Code of Practice), plus a third-party auditing regime.
- Whistleblower protections: universalized beyond SB 53's narrow catastrophic-risk scope. The authors endorse the AI Whistleblower Protection Act (Sen. Chuck Grassley), covering "substantial and specific" dangers even absent a legal violation.
- Information-sharing within and between governments (AISI–CAISI joint evaluations; EU AI Act Art. 78(5)), with antitrust safe harbors for Frontier Model Forum-style safety coordination.
- Flexible rules and definitions: if-then commitments; management-based regulation; and regulatory rather than statutory definitions of "frontier model," noting that SB 1047's $100M compute threshold was nearly obsoleted within months by DeepSeek.
- Assessments and evaluations: codify and fund CAISI with an expanded mandate, and subsidize a third-party evaluation ecosystem (METR, Apollo Research).
- Securing model weights and algorithmic secrets: voluntary physical and cybersecurity standards across the frontier supply chain, modeled on DoD's CMMC, with compliance tied to federal grants and contracts.
- Hiring and talent: a private-sector expert "reserve corps"; reform of the Intergovernmental Personnel Act; and raising pay at CAISI and the EU AI Office.
Preemption
The essay devotes a distinct section to the June 2025 state-AI-regulation moratorium, which was stripped from the reconciliation bill in the Senate. The authors call it "ill-advised," in part because of its effect on optionality: preempting state law and "replacing it with nothing" radically shrinks the regulatory option space, given how rarely Congress legislates. They argue that preemption should be narrow and iterative — adopted after a federal standard exists, not before — which they describe as "the only approach to preemption of state laws regulating an emerging technology that has ever been taken in the history of the United States." See State-Level AI Regulation and Techno-Federalism.
Objections addressed
The essay responds to several objections:
- "Giving the government a hammer" is answered with democratic oversight (congressional and White House oversight in the US; parliamentary oversight in the EU) rather than hamstringing regulators.
- On democratic legitimacy, the authors acknowledge a real tension between flexibility and legitimacy, but argue that failing to prepare raises the risk of undemocratic outcomes such as Aschenbrenner's "Manhattan Project."
- On concentration of power, the authors explain this as their reason for not recommending the expansion of emergency authorities such as the Defense Production Act; they cite "the ongoing dispute between the Pentagon and Anthropic" as illustrating emergency-authority abuse, and propose requiring government to use only law-following AI.
- "Private governance is all you need" (Dean Ball's Framework for Private Governance; Hadfield & Clark's Regulatory Markets) is answered with the argument that even these proposals require a competent government office to license and supervise private regulators, and that tort liability and the state's monopoly on enforcement cannot be replaced.
Framing and scope
The essay is a position, a programmatic governance argument, rather than evidence. The authors present its distinctive contribution as reframing the regulate/don't-regulate binary as a question of institutional readiness, and identifying a class of capacity-building policies they argue are nearly costless to innovation. They pitch it as "realistically achievable now" in both the US and EU, explicitly bracketing more ambitious proposals including international treaties, compute moratoria, Aschenbrenner's AGI Manhattan Project, and Buterin's d/acc.
Relationships
- instance-of: Radical Optionality (the named framework this essay introduces)
- supports: Regulating Under Uncertainty (radical optionality is a distinctive response to it); AI Whistleblowing; AI Pre-Release Vetting
- related: Regulating Under Uncertainty, State-Level AI Regulation, Techno-Federalism, Compute Governance, Law-Following AI, California SB 53, New York RAISE Act, EU General-Purpose AI Code of Practice (2025), California SB 1047 (VETOED), Claude Mythos Preview
- contradicts: pure "permissionless innovation" / market-only framings; the 2025 broad state-AI moratorium