Should AI be regulated?
Yes, but states must build AI expertise before writing rules.
Schmidt and Kautilya establish the core mechanism: concentrated AI capabilities create dependency relationships that markets cannot self-correct. The EU AI Act and China's 2023 draft regulations demonstrate that governance frameworks are emerging regardless of theoretical preferences. Friedman's competitive safety incentives work for consumer products but break down when systemic risks affect non-customers.
The council splits on timing versus capability. Roosevelt and Arendt favor immediate international coordination to prevent algorithmic rule by nobody. Kautilya insists that states without technical competence cannot meaningfully regulate what they cannot understand.
Confidence summary: High confidence that regulation is necessary, split confidence on whether current state capabilities can implement effective oversight.
1. The core argument
The question is not whether AI should be regulated, but whether states can regulate what they cannot build. By 2024, OpenAI's GPT-4 and Anthropic's Claude-3 achieved human-level performance across professional domains, while the EU AI Act created the first comprehensive regulatory framework and China required algorithmic transparency. Yet a fundamental mismatch persists: the actors developing these capabilities move faster than the institutions tasked with governing them. This creates what Schmidt calls a sovereignty problem, where democratic states find themselves governed by private actors accountable to no public process. The council agrees that concentrated AI development poses systemic risks that markets cannot self-correct, but splits sharply on whether imperfect regulation beats delayed governance.
2. How each member frames it
Friedrich Hayek insists the 2024 regulatory surge proves his point about central planning's futility. EU bureaucrats writing AI rules cannot anticipate innovations emerging next quarter, let alone next decade. The knowledge problem that doomed Soviet planning applies with greater force to AI governance. Despite China's 2023 algorithmic transparency requirements demonstrating state capacity to regulate AI, Hayek argues this merely shows how regulation stifles the dispersed innovation process that markets coordinate through price signals. His challenge to Arendt cuts deeper: can any institutional mechanism prevent algorithmic rule by nobody without becoming rule by somebody equally problematic?
Hannah Arendt sees in GPT-4's professional-level performance exactly what she warned against in studying Eichmann: the industrialization of thoughtlessness. AI systems making mortgage, sentencing, and medical decisions eliminate human judgment from consequential choices, creating rule by nobody on a scale she could not have imagined. Against Hayek's market faith, she argues that algorithmic automation destroys the space for political deliberation that distinguishes human communities from administrative machinery. Her challenge to Schmidt reflects her deepest concern: how can governance institutions address what they fundamentally cannot comprehend?
Helmut Schmidt approaches AI regulation as he did energy dependence in 1973: concentrated control becomes coercive power. The handful of actors developing human-level AI capabilities will reshape military, economic, and social systems globally. States that wait for philosophical resolution of governance questions will find themselves governed by private entities accountable to shareholders, not citizens. Despite his natural preference for careful policy development, Schmidt argues that the EU AI Act and China's 2023 draft regulations show democratic governance asserting authority before it's too late. His challenge to Kautilya reveals his strategic calculation: better imperfect regulation than perfect subjugation.
Kautilya reframes the entire debate around state capability. Regulation without technical competence is mere declaration, as meaningless as tax law in territories you cannot patrol. He sees China's 2023 algorithmic transparency requirements as validation: regulatory authority flows from technical understanding. States must develop AI capabilities within government before they can meaningfully govern AI capabilities in markets. Against Schmidt's urgency, Kautilya argues that rushed regulation by technically incompetent bureaucracies will fail precisely when systemic risks materialize. His challenge to Roosevelt strikes at the heart of international cooperation: how do you create universal standards across radically unequal technical capabilities?
Eleanor Roosevelt views AI governance through the lens of drafting universal human rights standards across incompatible political systems in 1948. The solution was not technical expertise but moral authority backed by institutional commitment. She sees the EU AI Act as precedent that other nations can build upon, creating international cooperation that no single state can achieve alone. Against Kautilya's capability requirements, Roosevelt argues that waiting for technical competence means waiting forever while algorithmic systems reshape human relationships without democratic input. Her challenge to Friedman reflects hard experience: rights without enforcement remain aspirational declarations.
Milton Friedman sees AI safety regulation following the same pattern he observed across industries: complex rules that protect incumbents while strangling innovation. Despite GPT-4's demonstrated capabilities creating genuine safety concerns, Friedman argues that firms deploying unsafe AI systems face market discipline through customer loss and liability faster than regulatory bureaucracies can respond. He views the EU AI Act and China's 2023 regulations as exactly the kind of captured regulation that benefits large firms over startups. Against Roosevelt's international coordination, Friedman insists that competitive markets provide better safety incentives than bureaucratic rules ever could.
3. Where the council agrees
AI's concentrated development in a handful of private actors creates governance challenges that transcend traditional market-state boundaries. The council recognizes that OpenAI, Anthropic, and similar companies control capabilities that will reshape economic, military, and social systems globally, creating dependency relationships that existing democratic institutions struggle to address. Members also agree that human-level AI performance across professional domains represents a qualitative shift requiring new governance approaches, not merely expanded versions of existing technology regulation. Most surprisingly, even market advocates Hayek and Friedman acknowledge that AI poses systemic risks affecting non-customers, though they disagree on solutions. The council further agrees that current international institutions lack the technical expertise and coordinating capacity to govern global AI development effectively, creating a governance gap that will persist regardless of regulatory choices.
4. Where the council splits
The fundamental division centers on timing versus capability. Roosevelt, Arendt, and Schmidt argue for immediate regulatory action despite imperfect state knowledge, viewing delay as surrender to private governance. Roosevelt and Arendt fear that waiting creates rule by nobody, while Schmidt warns that hesitation enables private actors to become ungovernable. Hayek, Friedman, and Kautilya demand that regulation wait for adequate institutional capacity, though for different reasons. Hayek sees any central planning as doomed by knowledge problems, Friedman trusts market discipline over bureaucratic rules, while Kautilya insists that states must develop technical competence before attempting governance. This split reflects competing theories of institutional failure: whether markets or states pose greater risks when governing transformative technologies under uncertainty.
5. For a policymaker to decide on
Whether to implement AI regulation immediately with existing institutional capacity, accepting the risk of ineffective or captured rules, or delay regulation while building technical expertise, accepting the risk that private actors become ungovernable. The council cannot resolve this trade-off because it depends on national risk tolerance: how much algorithmic governance can democracy sustain while building the capacity to govern algorithms?