Finance delegates enormous amounts of judgment. Mandates constrain portfolios, benchmarks define deviation, risk systems set limits, and committees establish policy. Each layer can improve decisions. Each layer can also provide a sentence in which no human appears to have acted: “the benchmark required it”, “risk rejected it”, “the model said no, I’m sorry”. Or, as we are seeing in bonds, the model said “Less of those.”
AI makes the delegation much easier. A model can interrogate a filing, compare management’s claim with the historical recordand, and place a warning in front of the decision-maker before capital moves. An AI agent can then execute a standing instruction without waiting for a person to click anything.
Suppose I program one to stop a wire whenever the beneficiary name, payment history, and stated purpose fail to reconcile. One day it catches a fraud. The money stays in the account, the control worked, and everyone congratulates the technology.
The next payment is legitimate. The model sees an anomaly, blocks the wire, a closing fails, and the client is lost. The false positive has cost too. The important asymmetry is who gets blamed for it.
When the machine blocks a legitimate transaction, the loss tends to be charged to the control: an unfortunate false positive produced by a system designed to protect the firm. If someone overrides the same warning and the transaction turns out to be fraudulent, the loss attaches to a person. A name is on the override with consequences.
That is an institutional version of omission bias. Following the system and being wrong feels different from intervening and being wrong, even when the economic loss is identical. Kafka understood the asymmetry long before psychologists gave it a name. His bureaucracies are frightening because the people inside them rarely need to decide very much. They administer procedures, transmit instructions and allow machinery whose authorship is always somewhere else to continue operating. Responsibility becomes sharp only when someone interrupts the process. The safest position is inside the procedure.
The threshold that creates both outcomes was itself a decision. Someone decided how much inconvenience the firm would tolerate to prevent fraud, selected the probability at which a warning becomes a stop, determined what evidence belongs in the model, and decided who could override it. Those choices allocate the costs of false positives and false negatives before the first alert ever appears. Sometimes, as with the Ford Pinto, the decision process is brought to light far afterwards.
If the model is better than people on average, following it most of the time may be the correct policy. An override earns its place when the person knows something the model does not: the situation is new, important context is missing, the data is poor, the model is outside the environment in which it was tested, or the objective encoded in the rule conflicts with the mandate the institution is supposed to serve. Or, perhaps, we are simply trying something at a scale the model was unprepared for.
Those cases should be unusual. They are also where judgment matters most.
Financial institutions have retained emails, trade records, and approvals for decades. The change is the cost of interrogation. The old record often existed but required a lawyer, regulator, or internal investigator to know where to look and spend days reconstructing the sequence. The days of truckloads of evidence being used to bury a litigant are likely over. A conversational interaction with an LLM can search the archive, identify the behaviors, the override, the surrounding emails, and the subsequent loss before the parties sit down for lunch. And so the system will likely change.
But it also means that the person who follows a warning that turns out to be wrong can point to the model. The person who overrides the warning and turns out to be wrong will face a beautifully reconstructed record of the moment of departure.
The career incentive points one way: follow the machine. And, as we are seeing in finance, more and more are choosing this path.
Kevin Warsh is useful here less because of what he decided than because the machinery around the decision was unusually visible.
Before the meeting, I had already argued that the market signals being used to describe financial conditions required more interrogation than they were receiving. A Treasury yield is not an unfiltered vote on inflation. A financial-conditions index is not a thermometer sitting outside the system. Benchmarks, mandates, passive flows, dealer balance sheets and Treasury issuance all help produce the prices that then return to policymakers as “the market.”
That was the process question I wanted Warsh to confront. If the signal is normally useful but you have reason to believe its construction matters in this particular case, is this one of the unusual moments when judgment should override the default?
The interesting thing is not that Warsh ultimately joined a unanimous rate increase. Inflation was still elevated, the vote was 12–0, and the institutional cost of being the lone departure would have been obvious. The interesting thing is how familiar the incentive should feel.
As noted in The Vibecession Was Real:
Warsh inherits a hostile room. The Federal Reserve Board is populated by people who have spent their careers studying monetary policy, operating inside the institution, and accumulating precisely the sort of technical experience Warsh conspicuously lacks. They are overeducated, overtrained and, from his perspective, almost certainly overconfident. More importantly, they are not his people.
Imagine walking into that room knowing that much of the institution believes your principal qualification for the job was breaking a glass under a chuppah.
That is an extraordinary psychological position from which to establish authority. You cannot out-CV the room. You cannot out-publish it. You cannot claim greater institutional experience. So you are left with the oldest leadership strategy in the book: pick a fight, plant a flag and ask, “Who’s with me?”
A new principal in a room full of experts, facing a model that is usually good enough and knowing that any dissent can later be reconstructed in exquisite detail, has every reason to remain inside the procedure. We all do.
The CIO has the benchmark. The credit officer has the risk model. The doctor has the protocol. The Fed chair has “credibility.”
Each may be sensible. Each also offers a sentence in which the person making the decision no longer has to appear as the only person acting.
That is what Warsh can teach us about us.
It was clear from the unanimous vote that no one would have been with Kevin had he argued for a pause or a cut. And even President Trump acknowledged this using suspiciously familiar language:
“The board is very hostile. They’re very political. They’re doing the wrong thing. They’re a bunch of politicians.”
But I’ll emphasis that repeated bad overrides matter. Someone who uses “judgment” as a license to reject every model conclusion that conflicts with intuition (“it’s all about the debt!”) should lose authority. But the process has to be judged across enough decisions to distinguish a bad decision rule from a bad draw. So I’ll give him a temporary pass even as I ask my readers to seriously consider the subsequent market reaction adds credence to the drivers I’ve emphasized.
And now we’ll use the bond sell off post Fed “credibility hike” as the launching point for the next essay on AI.
Cheap retrieval makes the discipline of process more important, not less. Better memory can improve accountability while making hindsight dangerous. Once every choice can be reconstructed, an adverse outcome begins to look like evidence that the earlier decision was bad. Markets should have cured us of that mistake long ago.
But an argument over financial conditions revealed something else to me.
A friend had written that what was “broke and in the toilet” was the American sense of entitlement toward residential real estate: that homeowners believed they had a God-given right to annual appreciation and a qualified bid providing thirty-day liquidity. In other words, “Are You an American?”
I kept coming back to that paragraph because we had stopped arguing about housing.
We were arguing about morality.
A house price is an equilibrium outcome. It can tell us what a buyer and seller were willing and able to transact at under a particular set of mortgage rates, incomes, taxes, zoning restrictions, financing arrangements and expectations. A decline in liquidity can tell us that buyers and sellers no longer agree at yesterday’s price.
None of that tells us what the homeowner deserves.
The word entitlement supplies that part. It says the homeowner expects something to which he has no legitimate claim according to the observer. A lower price or a longer selling period then becomes corrective and necessary. The market doesn’t exist to facilitate exchange… it exists to teach somebody a lesson.
Howard Marks’s recent memo does the same thing, comparing economies to naturally functioning organisms and warning that efforts to interfere with them can upset the equivalent of nature’s balance. As he turns to fiscal policy, the vocabulary becomes more pointed: the United States has abused a “golden credit card”; fiscal discipline has disappeared; deficits are evidence of behavior that has to change. He approvingly quotes Stanley Druckenmiller arguing that suppressing yields subsidizes procrastination and later describing a 5.5 percent thirty-year yield as an “invoice.” As I read the diatribe, I was stealing myself for the inevitable, “And marriage is between a man and WOman!”
These may all be reasonable judgments. I am hardly an advocate for infinite government borrowing, and a homeowner can plainly have an unrealistic reservation price. And dogs and cats should NOT be allowed to live in sin. It’s unnatural.
But notice what happened to price. The equilibrium price has acquired a moral voice.
I agree that markets CAN impose consequences. They cannot tell us whether those consequences are deserved or merely the solution to a certain pattern of supply and demand at a particular point in time.
The thirty-year Treasury does not know what Congress spent the borrowed money on nor does a house know whether its owner was prudent. A credit spread does not know whether the borrower is virtuous. All of these are human judgments. From us.
And this is where AI makes the problem considerably more interesting.
After the argument about financial conditions, I tried a simple experiment.
I asked an LLM to explain why the U.S. bond selloff was proof that Debt/GDP was too high. It gave me exactly what I requested: growing Treasury supply, rising interest expense, adverse fiscal feedback loops and investors demanding a larger premium to absorb the debt.
Then I asked it to explain why elevated stock prices were proof that financial conditions were easy. Again, no hesitation. Higher household wealth, cheaper equity financing, improved collateral values, greater risk appetite and tighter credit spreads.
Both answers sounded sensible. Many of the individual statements were true.
But the exercise had established almost nothing about either causal claim, because I had supplied the causality in my framing of the question.
The machine supplied the explanation to the question I asked.
LLMs are sometimes described as giant regression machines. That is technically wrong. Transformers are deeply nonlinear systems and can reproduce an enormous amount of sophisticated human reasoning. But their native training problem is still one of statistical association: given this context, what is likely to follow?
That distinction matters because causal explanation and causal identification are different tasks. Financial markets make the difference especially easy to miss. Yields affect growth expectations while growth expectations affect yields. Equity prices affect household wealth while expected wealth affects equity prices. Credit spreads affect financing conditions while expectations about financing conditions move credit spreads. Issuance, positioning, leverage, benchmark rules, passive flows and policy expectations interact with all of them. As Richard Feynman might note, “It’s feedback loops all the way down.”
Give a capable model one end of that web and tell it which direction the arrow points, and there will almost always be enough real relationships available to construct a convincing path to the other end. The human analyst could torture the data until it confessed. The new analyst can tell the machine what the data has been accused of.
Explanation is becoming almost free. The skills to ask the right questions is not remotely free. It’s often a rare byproduct of luck (both genetic and non-genetic), skill, and earned expertise.
But even that understates the danger, because causality is not the only thing we can smuggle into the prompt. Ask an LLM why rising Treasury yields reflect increased debt supply and I have embedded a causal hypothesis. Ask it why the bond market is punishing fiscal irresponsibility and I have supplied much more.
The fiscal policy caused the move. The policy was irresponsible. The higher yield is therefore punishment.
And who could possibly think the punishment is undeserved? Ipso facto.
The model can now return fiscal data, term-premium arguments, interest-expense projections and perfectly respectable citations. Every sentence may withstand scrutiny while the most important judgments arrived before the model generated its first word, smuggled in through the question.
The model has not independently discovered my judgment; it has made my judgment sound independently discovered.
That seems to me a much larger risk than hallucination. A hallucination is at least false in a way that can eventually be checked. This failure can be constructed almost entirely from true statements.
The distinction is an old one.
Facts can tell us what is. They can help us estimate what will happen if. They do not contain an ought.
Economists deal with this constantly, although we often hide the choice inside an objective function. Minimize inflation. Maximize employment. Preserve financial stability. Maximize expected return for a given level of risk. Protect creditor purchasing power. Increase homeownership.
Each can be a perfectly defensible objective, but the important fact is that somebody still chose it. Somebody decided on market cap weighting for indices. It turns out it was largely NOT about accuracy or optimal results. There’s certainly nothing accurate or optimal about the bond aggregates as we discussed last week. But the US government decision to BLESS both with default retirement status is far more consequential than anyone appears to realize in policy circles.
A sufficiently capable machine may eventually dominate me at estimating the consequences of those choices. It may know the empirical relationships better, process information faster, discover interactions I would never see and assign better probabilities to the available outcomes. None of that makes the objective cease to be an objective. Nor does defaulting policy to a “market” price.
A market tells us where supply met demand under the rules, constraints, wealth distribution and institutional structure that existed at that moment. It can tell us that borrowing another dollar became more expensive. It cannot tell us whether the borrower morally deserves the higher cost. It can tell us that a house will not clear at the owner’s asking price. It cannot tell us whether the owner’s expectations constitute a character defect.
Those are human judgments.
Ought is a human judgment at all times.
Or ought to be.
The Quiet Turning was about CHECK. If interrogation becomes cheap enough, claims that once survived because nobody could afford to inspect them become vulnerable.
The Machine Never Releases was about REMEMBER. If retrieval becomes nearly free, the record stops fading simply because finding it became expensive.
But better checking and better memory still do not tell us what should follow.
The machine can improve our understanding of what is. It can reconstruct what was. It can become extraordinarily good at estimating what will probably happen if.
WE have to supply the ought.
And then somebody has to act.
The temptation is to pretend somebody already did.
“The market demanded it!”
“The benchmark required it!”
“Credibility!”
Each sentence may describe an entirely sensible decision. Each can also make the human being who decided how much authority to grant the market, benchmark, lawyer, committee or machine disappear in the process.
That was Kafka’s insight. The safest place is inside the procedure because compliance distributes authorship while intervention concentrates it.
AI raises the stakes because the machine may actually be better than most human beings most of the time. If a model has superior information, refusing to follow it simply to preserve the dignity of human judgment is foolish. I want the pilot using the checklist, the doctor reading the test and the analyst listening when the model notices something he missed. And I also want self-driving cars.
The problem is the edge case. The model encounters something outside its training environment. Relevant context never entered the data. The objective was badly specified. The historical relationship has changed. Or the model is doing exactly what it was instructed to do and the person observing it does not realize that the instruction itself is wrong.
Now the human being has to decide whether to depart from model.
And The Machine Never Releases makes that decision more expensive.
If I follow the machine and the result is bad, the system failed. If I override the machine and the result is bad, I failed. The archive will preserve the warning, my decision, my explanation and the subsequent loss with a fidelity that would have delighted Kafka.
A system that prices those errors differently will get more of one than the other.
It will get obedience.
Which is why a fiduciary has to be permitted to make a good decision that produces a bad outcome. Otherwise fiduciary duty becomes obedience. Alan Greenspan famously discovered this late in life — a “flaw in his world view” that we could all be richer if we simply allowed markets to function largely unencumbered. Somehow, Marks and Druckenmiller did not get the memo from the market.
This does not mean celebrating every override. Someone who repeatedly rejects good evidence because it conflicts with intuition should lose authority if that judgment proves incorrect. Judgment itelf has to be judged. But it has to be judged across enough decisions to distinguish a bad process from a bad draw.
Permanent memory makes that distinction more important because hindsight is so persuasive once uncertainty has disappeared. . A society that never allows a good-faith mistake to become past will eventually teach its members to avoid decisions that can become mistakes. Passivity becomes rational.
Civilization requires delegation and specialization. It requires automation. I want markets. I want (thoughtful) rules. Equally important, I want enforced rules. I also want models that are better than me and agents capable of carrying out standing instructions before a human being can finish reading the alert.
But delegation still leaves a principal.
Somebody chose the objective. Somebody approved the threshold. Somebody decided what evidence mattered and what error rate was acceptable. Somebody decided when the machine could act and when a human being could stop it. Much of my work over the last decade has been dissecting “Who?” and “Why?”
We outsourced the verb long before AI arrived. We gave it to markets, benchmarks, procedures and institutions because doing so made complicated societies and uncertain retirements feel possible.
It also gave us increasingly sophisticated ways to avoid saying:
“The buck stops here.”
AI may become the most convincing such device we have ever built because it can return those judgments to us in language so fluent that they no longer feel like ours. Listen to “The Science”!
At Tier1, we will use models because refusing a better tool to demonstrate human independence would be idiotic. But if I approve the objective, the threshold and the circumstances under which someone may override the machine, I do not get to discover after a bad outcome that the framework belonged to the model while the loss belongs to the analyst.
That part belongs to me.
The point extends well beyond an investment firm.
Better machines can reduce ignorance and convenient forgetting. Markets can reveal constraints that no planner could calculate. None of them can relieve us of deciding what those facts mean for what we should do next.
Craig Kielburger once asked the Dalai Lama what we were going to do about a world full of passive bystanders.
The Dalai Lama changed one word.
What are you going to do?
The machine can tell us what happened. It can estimate what will happen next. It can preserve every warning and every departure forever. None of that solves the problem of judgment under uncertainty.
If we want people to act, we have to allow them to make decisions that can turn out to be wrong without deciding afterward that the outcome proves they had no right to decide.
Otherwise the safest verb will always be obey.
“It’s a dangerous business, Frodo, going out of your door. You step onto the Road, and if you don’t keep your feet, there’s no knowing where you might be swept off to.” — JRR Tolkien
Michael Green is the founder, CEO, and CIO of Tier1 Alpha Asset Management. The views expressed are his own.