Can AI systems legitimately resolve wicked policy problems?

Peter Levine, writing in National Civic Review (March 2026), tests whether AI can resolve "wicked problems" — Rittel and Webber's 1973 term for policy problems that have no definitive formulation, no stopping rule, and where "choices are not true or false, but good or bad." Levine's central move is to reframe what AI is: not "like a human brain" but "like a social institution" — comparable to medicine, labor markets, or integrated assessment models for climate policy — because all of these "aggregate vast amounts of information and huge numbers of decisions" into outputs no single human produced or can fully audit. On that model, AI is "not fundamentally different from a given social institution, such as a scientific discipline, a market, a body of law, or a democracy," and so could in principle be asked for, and give, answers to wicked problems.

The catch, in Levine's own terms, is that "AI's rules are different from those in law, democracy, or science" — and he is "biased to think that its rules are worse." He tests this by asking ChatGPT and Claude directly whether AI can resolve wicked problems; both decline — ChatGPT because such problems "are value-laden" and deciding "whose values count, how to weigh them, and when to revise them is a normative, political act, not a computational one," Claude because "stakeholders can't even agree on what the problem actually is." Levine then surveys three rival accounts of how people decide value questions — subjective preference, a computable right answer such as utilitarianism, and strict subjectivism — and finds each too weak to hand the decision to a machine: preferences can be bad ones, a right-answer calculus presumes moral reasoning is algorithmic when it looks "holistic" instead (judging whole objects like "a school or a market," not separable variables), and pure subjectivism undercuts itself. His conclusion is not that AI lacks the capability to generate answers, but that today's models refuse "because their authors have chosen — so far — to tell them not to," a policy choice, not a technical limit.

This cuts against a purely capability-based skepticism of AI and normative reasoning. Can AI distinguish which differences actually matter? argues AI's limit is structural: it pattern-matches rather than makes the qualitative judgments of relevance that expert observation requires. Levine's holism point — that moral reasoning judges whole objects rather than decomposable variables — makes a parallel case specifically for value judgments. And where Should AI alignment target preferences or social role norms? argues alignment should track role-appropriate normative standards rather than aggregated preferences, Levine arrives at a related worry about legitimacy from the opposite direction: even an institutional AI that could compute an answer would still lack standing to settle a political act like weighing whose values count. His "AI as institution, not brain" framing also sits near Do classical knowledge definitions apply to AI systems?, though Levine's angle is legitimacy for value choices rather than the epistemology of knowledge claims.

The essay is explicitly tentative ("I am still tentatively using the following model") and offers no empirical test of whether AI's answers to wicked problems would in fact be worse than human institutional answers — only Levine's statement that he is "biased to think" so. It also does not specify what "AI's rules" being worse actually consists of beyond the refusal behavior trained into current chatbots, nor what follows once a future model's authors choose differently. The claim the piece actually supports is narrower than "AI can't solve wicked problems": current refusals are policy, not incapacity, and the real risk is adoption by default — people accepting AI's answers to value-laden problems because "our own responses are not definitively better and because it responds instantly at low cost," regardless of whether those answers are actually better.