What this covers
Austan Goolsbee, president of the Federal Reserve Bank of Chicago, argues for the irreplaceable role of human judgment in steering monetary policy—not as a counterweight to data, but as a prerequisite for reading the economy's underlying structure. The conversation ranges across his diagnosis of recent inflation, the limits of mechanical rules and historical precedent, the dangers of over-indexing on artificial intelligence, and the architectural choices the Federal Reserve must make to avoid institutional groupthink. Throughout, Goolsbee emphasizes a core claim: that policymakers cannot navigate structural change or supply shocks by leaning on accounting identities or patterns learned from demand-driven cycles. The 2023 disinflation—a near-collapse in price growth without recession—serves as his central exhibit: it refutes the historical priors embedded in many rule-based frameworks and illustrates the cost of treating past business cycles as reliable blueprints.
The discussion moves into several territories. Goolsbee examines why the post-pandemic inflation puzzle required diagnosis rather than formula, distinguishing supply shocks from demand shocks and showing where standard demand-centric machinery fails. He considers the structural vulnerabilities in financial architecture—the 500-year problem of money-like deposits redeemable on demand, the tension between overseen banking and shadow banking, and why stablecoins pose systemic risk if they attach to the payment system. On artificial intelligence, he contends that training sample limitation would have blinded an AI system to the 2023 supply shock just as hallucination makes it unsuitable for domains where error is unforgivable. He also addresses productivity puzzles—why construction has negative productivity growth for decades, why housing costs have compounded upward for fifteen years, and whether current AI productivity gains will diffuse across the economy or peter out. Finally, he defends the Federal Reserve's decentralized structure, arguing that diversity across regional Reserve Banks resists the monoculture risks that concentrate power in a single decision-making center.
Goolsbee argues that good monetary policy requires human judgment and a micro-grounded supply-vs-demand diagnosis rather than mechanical rules, historical training samples, or single models—because the most dangerous errors come from over-indexing on past patterns (whether monetarist rules or AI trained on the past) when supply shocks and structural change are driving the economy.
- The 2023 disinflation without recession shows rules and historical priors premised on demand-driven cycles can mislead.
- Accounting identities (MV=PY) have no economic content and shouldn't be the basis of pre-committed policy rules.
- AI/LLMs are 'only as good as the training sample' and would have missed the 2023 positive supply shock, just as human judgment was needed.
Demand-centric monetary policy framework produces wrong answers during major supply-side shocks.
- Before drawing any policy conclusion you must diagnose whether a development is a supply shock or a demand shock; central banking machinery is overwhelmingly oriented around demand, so during major supply-side developments (labor supply, supply chain, productivity) the standard demand-centric approach will produce wrong answers.
“before you can conclude anything, you’ve got to get a taste of, is this a supply shock or is this a demand shock?”
- Policymakers should not design rules off accounting identities like MV=PY because identities contain no economic content; a rule premised on the assumption that shocks resemble previous business cycles (demand-driven) is not robust to changed conditions and should not be pre-committed to.
“I don’t like making policy off accounting identities. There’s no economic content in accounting identity.”
- A purely demand-driven explanation of post-pandemic inflation faces three puzzles: inflation soared while unemployment was above 6 percent (well above the natural rate); similar-magnitude inflation hit many countries (including domestically-driven economies like Nigeria and India) that lacked comparable stimulus; and inflation did not fall when stimulus rolled off but did fall once supply chains healed in 2023.
“why did inflation begin soaring in the US when the unemployment rate is over 6 percent?”
- In 2023 inflation fell nearly as much as it ever has in a single year without a recession, contradicting the historically-rooted view (e.g., Bloomberg economists' 100 percent recession prediction) that significant disinflation always requires serious recession—demonstrating the danger of rules implicitly premised on demand-driven cycles.
“There had never been a drop of inflation of a significant degree without a very serious recession. Yet in 2023, there was. Inflation fell almost as much as it ever fell in one year without a recession.”
Risky money-like instruments attached to payment systems create financial stability crises.
- Volcker's principle from the 2007-08 crisis is that anything too risky must stay outside the shadow of rescue—if it's risky it can't get rescued—and the critical danger arises when risky instruments like stablecoins attach to the payment system, because in crises society must ensure people can still make payments (rent, mortgage, groceries).
“Anything that’s too risky has got to stay outside the shadow of rescue, if you want to think of it that way. If it’s risky, it can’t get it.”
- Any entity that creates money-like deposits redeemable on demand will end in bank runs and tears unless it has either deposit insurance or serious restrictions on backing assets and permissible activities—a lesson supported by 500 years of financial history.
“there better be some deposit insurance, or... pretty serious restrictions on the assets backing that... or else that thing ends in bank runs and tears. We have 500 years of financial history that tell us that.”
- Financial stability faces a core conundrum: an official banking sector under central bank oversight coexists with a growing shadow banking sector prone to runs that the central bank cannot easily oversee, raising the unresolved question of the lender-of-last-resort role where the central bank doesn't operate.
“there is an official banking sector that we have oversight of. Then there’s a shadow banking sector, which is getting bigger and bigger every year and can be prone to runs”
- The US free banking era after the Second Bank of the United States was ended—featuring the Panics of 1847, 1857, and near-panic of 1860 with massive bank failures where anyone could start a bank, print money supposedly backed by treasuries, and skip town—provides a cautionary precedent against allowing risky stablecoins to become part of the payment system.
“If you go look at the free banking era, it is an era of massive numbers of bank failures and recurrent financial crises. The Panic of 1847, the Panic of 1857, and close to panic in 1860”
AI productivity gains will diffuse gradually across sectors following the pattern of prior general-purpose technologies.
- If AI is a general-purpose technology like computers, electricity, or telephones, its productivity gains will diffuse over decades from the producing sector to using sectors—as seen with computers, where productivity appeared first in computer production, then years later in computer-using sectors, then later still in firms like Walmart adopting IT—and Chicago Fed work shows the current productivity surge is concentrated in tech/AI-intensive industries, consistent with this diffusion pattern.
“If AI is a general-purpose technology like computers, like electricity, like telephones — that sort of thing — those played out over decades.”
- To the extent AI improvement comes from bigger data sets and more compute rather than new algorithms or theory, diminishing returns will kick in quickly—much as self-driving cars improved rapidly 15 years ago before improvement slowed dramatically rather than going to zero—so grand AI adoption forecasts that extrapolate current growth rates are likely overstated.
“data sets and computing power have diminishing returns that will kick in pretty quickly.”
- There is a high danger that anticipation of future AI bounty drives surging equity values and AI-infrastructure investment, producing capacity-constraint and over-capacity problems, while wealth effects from high equity valuations can boost present consumption—both of which can cause overheating if monetary policy is not careful.
“if equity values surge and business investment surges in AI infrastructure-type capital investment on the anticipation of the bounty that is to come, I think there’s a high danger that you encounter the over-capacity, the capacity constraint problem”
Long-run housing relative price inflation and construction productivity stagnation drive persistent housing despair.
- Research with Chad Syverson documents that the construction industry has had negative productivity growth over long periods—at best zero and mostly getting worse at building—a genuine puzzle that may be partly tied to land-use and other regulation forcing inefficient scale, since it appears even in rural areas with less land-use regulation.
“negative productivity growth in the construction industry over long periods of time, which is itself a puzzle.”
- The relative price of housing has been rising about 4-5 percent per year for a decade and a half—in the 12 years before COVID house price inflation ran 3.5-4 percent annually while goods prices deflated about 1 percent annually—and this long-run compounding, not just a recent bubble, explains widespread housing-cost despair across both cities and rural areas.
“house price inflation was 3.5 percent or 4 percent a year, and goods price inflation was actually deflation of around 1 percent a year. The relative price of housing has been rising 4 percent, 5 percent per year for a decade and a half.”
- The old relationship between M2 and price inflation broke down because financial innovation (electronic payments, credit/debit cards replacing cash and checks) radically changed the velocity of money; when both M and V move simultaneously, over-indexing on M=PY interpretations becomes misleading.
“It wasn’t just M equals PY. It was M times V. If M is moving around, and V is moving around at the same time, you’re going to get a little mixed-up interpreting to over-indexing on that theory.”
- Eric Budish's theory paper shows blockchain relies on collective agreement about transaction history, creating an attack vector: controlling 50 percent plus one of the instantaneous mining flow lets an attacker rewrite history and snipe-attack a large stock of crypto, illustrating a structural vulnerability of non-centrally-overseen crypto.
“if you could get 50 percent plus 1 of the miners... you could get 51 percent of the instantaneous flow, and that would allow you to snipe-attack a giant stock of crypto.”
- The post-2008 efforts to raise capital ratios at the largest official banks worked: unlike 2008-09 when big banks were the centerpiece of the problem, in the March 2023 Silicon Valley Bank events the largest official banks were not the source of critical capital deficiencies, suggesting capital regulation was effective rather than merely pushing risk to shadow banks.
“The fact that in March of ’23, the big banks were not the problem, I think a lot has to do with all of the efforts that were put in place to raise capital at those larger banks.”
- Economists cannot fully explain why money matters distinct from T-bills given interest on reserves, but the market behaves as if there is an unexplained 'yield to safety'; this becomes more relevant when many imperfect money-like substitutes (stablecoins, credit/debit instruments) proliferate, which cautions against returning to simple M2 thinking.
“there’s some yield to safety that we don’t quite understand, I’d say, and it becomes more relevant perhaps in a world where people are creating money-like deposits all over the place.”
- Goolsbee beat Ted Cruz in debate because Cruz's strength was pre-planned plotting and traps while his weakness was thinking on his feet when angered; teasing and joking to the judge would make Cruz red in the face and lose composure.
“His weakness was thinking on his feet, especially if he got mad, so I would, to the judge, start teasing.”
- The Federal Reserve performs five basic functions: monetary policy (the 'opposable thumb'), directing the US payment system (wire transfers, ACH, FedNow), bank supervision and regulation (the largest in employees), financial services for member banks (e.g., cash for ATMs), and community/regional economic development with local industry expertise.
“It’s five basic functions, one of which is monetary policy. I consider that the opposable thumb.”
- The MBA is a well-functioning market and valuable degree: judged either by market demand or content of skills taught, with competitive pressure operating mainly on the opportunity-cost side through shorter one-year or 18-month specialized master's programs in finance and management.
“I think it’s a pretty well-functioning market. It’s a valuable degree, and it teaches a lot of skills.”
AI systems hallucinate and cannot replace human judgment in domains where errors are costly.
- An LLM is only as good as its training sample, so an AI replacing the FOMC in 2023 would have wrongly prescribed jacking up rates to force a recession to bring inflation down, because it could not recognize the ongoing positive supply shock (rising labor force participation, more flexible labor markets, disabled workers and women re-entering)—the same limitation makes AI prone to hallucination as a central banker.
“you’re only as good as the training sample. The training sample would’ve said, as you went into 2023, inflation’s way too high... The LLM would’ve said, ‘It’s a lead-pipe cinch guarantee. You need to jack up the interest rate and have a massive recession.’”
- The hallucination problem is extremely damaging for applications where making a mistake is an unforgivable error; AI's greatest near-term opportunities are in domains without a single correct answer, and merely upgrading to higher-priced model versions does not solve hallucination—developers must change the penalty function, not just add data.
“there are some things for which making a mistake is an unforgivable error... You’ve got to change the penalty function in that thing.”
The Fed's decentralized regional structure prevents groupthink and preserves diverse economic perspectives.
- The Fed's hybrid federal structure—12 regional Reserve Bank presidents (non-political appointees chosen by regional boards) sitting alongside 7 political governors on the FOMC—is a durable design feature that ensures monetary policy reflects perspectives beyond New York and Washington and guards against the groupthink that afflicts monocultures, so consolidating research staffs would be a mistake.
“we need to let some flowers bloom because monocultures are prone to groupthink.”
- The most useful microeconomics for managing an institution like a Reserve Bank are the theory of delegation and the distinction between cost, opportunity cost, and marginal cost—activities with a marginal cost near zero (like sharing information or opening meetings, since knowledge is a public good) can be expanded freely for strong benefits.
“what is a cost, what is an opportunity cost, and what is a marginal cost, are some of the most important microeconomic ideas that I think we can apply to management.”