What this covers

John Kay and Mervyn King sit down with host Russ Roberts to challenge the foundations of how economists think about decision-making. The conversation rests on their claim that most important choices happen in large worlds of radical uncertainty—situations so complex and novel that probability-based models cannot capture them—yet economists have built their entire apparatus around optimizing expected utility in calculable, simplified settings. Against this approach, Kay and King propose a different method: asking "What Is Going On Here?" and reasoning through the specific situation rather than trusting precise numerical outputs from models.

The discussion moves across finance, epidemiology, and business practice to show how this gap between theory and reality produces concrete harms. Kay examines how value-at-risk models failed catastrophically because they drew risk estimates from historical data that could never contain the extreme events they were designed to protect against. King dissects spurious precision in pandemic forecasting—where projections of specific death tolls create false confidence and obscure genuine ignorance. Both speakers revisit Milton Friedman's influential as-if hypothesis, showing how economists forgot it was a modeling device and began treating it as a description of human behavior. They also rehabilitate what behavioral economists labeled cognitive biases, arguing instead that these are evolutionary adaptations to an uncertain world. Throughout, they draw on the older distinction between Risk vs Uncertainty developed by John Maynard Keynes and Frank Knight, and defend satisficing—finding solutions good enough for the problem at hand—over optimization as the realistic model of how human beings actually decide.

Sharpest takeaway

Kay and King argue that most consequential decisions occur under 'radical uncertainty' that cannot be quantified, so the economist's project of optimizing expected utility and producing precise model-based numbers is misguided; the better response is to ask 'what is going on here?' and embrace abductive reasoning.

  • Probability-based models work only in 'small worlds' where information is complete, not in the 'large worlds' of real decisions
  • So-called cognitive biases are largely evolutionary adaptations that have made humans the dominant species
  • Spurious precision (in epidemiology, cost-benefit analysis, value-at-risk) deflects responsibility and hides genuine ignorance

The argument · threads4 threads · 37 claims
0.85

Economic models are valuable teaching puzzles but fail when mistaken for descriptions of actual human behavior and world dynamics.

4 pointscentrality 5/5
  • Milton Friedman's 'as-if' methodology fails because any model prediction contains many assumptions — both the maximizing behavior and the model's structure — so when a prediction is falsified you can never identify whether the as-if assumption or some other part of the model is wrong; in practice economics has forecast woefully, predicting nothing significant and then dismissing it as an unexplained shock.

    “you can never easily identify which bit of the prediction is false, why the prediction is falsified. Is it because the as-if assumption is wrong or is it because part of the model itself is wrong?”

  • We should not expect to predict the economy because we live in a world of radical uncertainty where the laws governing behavior are not stationary — they are changing all the time — so the Friedman test of predictive success is one economics inevitably fails.

    “the laws that are governing behavior are not stationary. They are changing all the time. So we shouldn't expect to be able to predict.”

  • Optimizing models like Ricardo's comparative advantage and Akerlof's market for lemons are valuable as puzzles that yield genuine insights, but they are not descriptions of the world; the great danger is economists believing such models apply to the whole panoply of human life rather than to the narrow small worlds where the axioms hold.

    “We give examples in models where the assumption of optimizing behavior gives real insights.”

  • Expected utility maximization worked well analyzing a world of certainty as a natural extension of preference orderings, but extending it to uncertainty was a non-trivial departure that economists treated as mere relabeling of commodities or constraints, ignoring that radical uncertainty changes the very nature of the problem — which is what Knight and Keynes recognized.

    “it was when the extension was made to the world of uncertainty that I think we got into serious trouble”

0.78

Radical uncertainty makes precise probabilistic forecasts and calibrated numerical targets dangerously misleading.

3 pointscentrality 5/5
  • Most real-world uncertainty is 'radical uncertainty' — situations where you know something but not enough to assign meaningful probabilities; knowing pandemics could occur did not let anyone assign a probability to a specific virus emerging from Wuhan in December 2019.

    “our view of radical uncertainty is that it's uncertainty that you cannot easily quantify”

  • Calibrating precise bank capital and liquidity requirements produces arbitrary numbers, and the danger is that once banks meet those numbers everyone stops worrying regardless of underlying circumstances; better regulation runs many different kinds of stress to see which banks are vulnerable to which shocks rather than pretending to know the future stress.

    “as long as the banks meet those arbitrary numbers, whatever the underlying circumstance is, then they're fine and people stop worrying about it”

  • Producing precise numerical forecasts (like '2.2 million deaths' or '550,000') is not 'doing the best we can' but pretending to know more than we do, which leads decision-makers to act on false assumptions and lets politicians and experts deflect responsibility onto each other.

    “It's not doing the best we can. It's pretending that we know more than, in fact, we do.”

0.71

Decision-makers must ask what is actually happening rather than delegate judgment to algorithmic optimization or false mathematical certainty.

4 pointscentrality 4/5
  • Before the 2008 crisis, when Bank of England staff asked financial firms about the really big risks, firms replied they didn't worry about them because a really big risk 'will be your problem, not ours' — illustrating how bailout expectations shifted catastrophic risk onto the state.

    “they would say, 'Oh, we don't worry about that, because if it was a really big risk, that will be your problem, not ours.'”

  • The most important thing when confronting a decision under uncertainty is to ask the simple question 'What is going on here?', because most big decisions are unique one-offs that require thinking through the actual situation rather than relying on precise model outputs.

    “the most important thing to do when confronted with a decision under uncertainty, which is to ask the very simple question : What is going on here?”

  • When economists are appointed to serious public policy positions, they do not go around saying 'I maximized expected utility this morning using a model'; they approach decisions in a very different way, which shows that sensible people do not actually optimize and that economics needs to broaden its horizons.

    “whenever economists are appointed to serious public policy positions, they do not go around saying, 'I maximized expected utility this morning using a model and this is the result.'”

  • Delegating decisions to computers because they outperform humans is dangerous since it applies only to small worlds with complete information; the great secret of the human race is not speed but the ability to make leaps of imagination and adapt to ill-defined problems that may have no unique answer, as seen in countries struggling to figure out 'what is going on' with COVID-19.

    “what we are very good at are making leaps of imagination, dealing with problems that are ill-defined, which don't necessarily have a unique answer”

0.56

Epidemiological and economic models reveal qualitative patterns but cannot produce reliable quantitative predictions due to poorly understood human behavior and non-stationary laws.

2 pointscentrality 4/5
  • Epidemiological models are valuable for revealing the general shape of a pandemic (slow start, sudden acceleration, peak, decline, possible later peaks) which helps you reason about whether health services can cope, but they are not good at producing precise numbers because their parameters reflect poorly-understood human behavior; even their originator Roy Anderson said they are useful for teaching but not for quantitative prediction.

    “Roy Anderson said, 'These models,' he said, 'They're incredibly useful for teaching purposes, but they're not helpful in making quantitative predictions.'”

  • The dream of completing markets — buying and selling contingent claims for every conceivable future event — is an astonishing delusion, because for so many possible future events there are and never will be markets (there is no market in 2030 London-New York airline tickets), so creating ever more derivative instruments does not complete the Arrow-Debreu set and cannot solve the problems Keynes and Knight identified.

    “there are clearly no markets for them at all and never will be”