On July 4th, The Economist ran a guest essay by Carson Block, the short-seller who built Muddy Waters Research. The argument is compact enough to state in a paragraph, and it has been sitting in the back of my head since.
AI displaces highly paid knowledge workers. They stop paying into their retirement accounts and, in Block’s phrasing, "workers won’t just stop paying in, they will need to withdraw funds." The withdrawals hit passive index funds, which sell every constituent in proportion to its weight, so the selling lands hardest on the most concentrated names. Then the line that made me go back: "Ironically, the companies with the highest multiples, which present the greatest risk from net outflows, are largely those that make up the AI stack." The AI trade gets clobbered by the thing AI does.
Five months earlier, Andrew Yang argued the opposite direction from the same world. Companies will cut knowledge workers because the market makes them: the stock market will reward you if you cut headcount and punish you if you don’t, and an investor he quotes puts it more plainly still, sell anything that consists of people sitting at a desk looking at a computer.
Put the two in sequence and they close a circle. Equity prices are the standing incentive to displace, and displacement is what breaks equity prices. Yang is telling you to buy the companies doing the cutting. Block is telling you those same companies are where the damage lands. Neither man mentions the other, and the two mechanisms cannot both run to completion, because the second one eats the first one’s fuel.
I have no ability to forecast whether there will be a crash. But a loop that consumes its own precondition has to break somewhere, and asking where it breaks turns a prophecy into a mechanism you can check.
What is actually new here
Yang is the useful control, because he made the same argument eight years ago about different people. His "normal American" in 2018 had a net worth around $36,000, under $500 in flexible savings, and almost nothing in the stock market. The person Yang was writing about owned no equities.
Every automation panic before this one described the displacement of people who owned nothing. Weavers, switchboard operators, and in Yang’s version cashiers and truck drivers. Their unemployment was a demand story and a social story, and it stayed in those lanes, because no financial channel ran from their lost wages to asset prices. They were never the bid.
Block’s displaced worker is the bid. The top 10% of American households hold over 87% of corporate equities, and the top decile accounted for 49.2% of all US consumer spending in Q2 2025. The group being displaced, the group doing the spending and the group owning the assets have converged into one group. That is a fact about the present rather than a forecast, and it is the only part of Block’s essay that would survive if every one of his numbers turned out to be wrong.
The chain, and where it thins
Six links. AI displaces highly paid cognitive workers. Those workers are roughly half of US consumption. Their contributions stop and withdrawals begin, flipping net flows negative. Passive funds sell the index to meet redemptions. The price impact concentrates on the mega-caps, because demand for the largest stocks is the most inelastic. And those mega-caps are the AI stack.
Some of that is solid. Passive selling proportionally to meet redemptions is mechanically true, and the concentration underneath it is real: the top ten names reached 40.7% of S&P 500 weight at the end of 2025, against about 27% at the dot-com peak, with passive now above half of US fund assets. Block’s description of the index as "no longer a diversified basket but a concentrated volatility trap" is fair. The weight sits on the other three links.
A multiplier that is not in the literature
The mechanism needs outflows to move prices violently, and Block reaches for a number: multipliers "possibly as high as a multiple of 100 for the largest firms." A dollar of selling removing a hundred dollars of market value.
That number is not in the research. Gabaix and Koijen’s inelastic markets hypothesis puts it around 5, and that is for the aggregate market rather than any single stock. Haddad, Huebner and Loualiche, in the American Economic Review, put the representative micro multiplier near 2. The widest published firm-level range is Davis, Kargar and Li at 0.3 to 15. Block’s 100 is roughly seven times the highest published estimate.
The direction does have support. Micro multipliers rise with firm size, and twenty years of passive growth has made demand for individual stocks about 11% more inelastic. So the mega-cap asymmetry is defensible. The figure attached to it is a practitioner’s extrapolation wearing the clothes of a research finding.
We have already run this experiment, twice
The bigger problem is that the most dramatic step has already been tested, and it did not do what the theory says. Block’s trigger is specific: "when the professional class begins to draw down its accounts to service mortgages, the resulting market impact will be sudden and violent." That requires mass job loss to convert into mass retirement-account liquidation.
In 2008 it did not. Balances fell hard, but almost entirely from market losses rather than behaviour. In 2020 it did not either, despite the CARES Act making penalty-free withdrawals of up to $100,000 available. Most savers did not take it. In both episodes the largest swing ran the opposite way, because suspending required minimum distributions reduced outflows. What households did, twice, was cut their contributions and ride it out.
Block has a rebuttal, though he does not make it: those shocks were cyclical, whereas permanent obsolescence would differ. I find that genuinely plausible. But it is an assumption rather than evidence, and it is the one the entire violence of the scenario rests on. If displaced analysts in 2029 behave the way unemployed households behaved in 2009, the flow link never fires and there is nothing left to transmit.
Why this is an American mechanism
From where I sit in Helsinki, the striking thing is that the chain does not translate. Finnish earnings-related pensions are collective and buffered, with contributions mandatory and employment-linked, the TyEL rate fixed at 24.4%. Mass unemployment would cut the contribution inflow. What it would not do is generate individual redemption demands, because nobody can ring Ilmarinen and ask for their balance. No fund is forced to sell into a falling market.
Norway is the cleanest contrast. The Government Pension Fund Global holds over $2.1trn with no redemption liability, making it a structurally stabilising buyer of exactly the assets Block expects to be dumped. So the concentration exposure is shared, and the forced-selling mechanism is not. Block’s loop is a defined-contribution loop, which makes it a claim about institutional design rather than a claim about AI. The country that pioneered individual retirement accounts built the transmission channel into its own financial system. The ones that kept pensions collective did not.
Unless there is no loop at all
The strongest objection is that the first link never fires, in which case there is no loop, only an expensive capex cycle. Of more than 1.2 million announced US job cuts in 2025, JPMorgan counted about 55,000 that cited AI: under 5% of cuts, and roughly 0.03% of total employment. Block names the two counter-arguments, Jevons and scepticism about the adoption curve, and disposes of both in four words: "Both will be proved wrong."
Which is where Yang earns his place as a calibration instrument rather than a punchline. He leaned on Frey and Osborne’s estimate that 47% of US jobs were at high risk of computerisation; the OECD’s task-based redo got 9%, and the gap was methodological and identifiable in advance. His most confident specific was 2 to 3 million driving jobs lost in 10 to 15 years, and there are now fewer than 50 driver-out trucks running against an 82,000-driver shortage. His least dramatic prediction, the quiet clerical grind, is simply happening.
Block’s "15% of jobs in America’s broader knowledge economy" within three to four years is the same species of estimate as the 47%, and deserves the same discount. So does the loop I have just drawn, which is an occupation-level displacement story wearing financial plumbing.
What is left when the numbers go
Block’s ending is the part I would keep if I could keep only one, and it is a strange thing for a short-seller to write: "stabilising the financial markets will be the easy bit for governments. They will struggle far more to manage the reordering of society." Central banks have a tested 2008 playbook for liquidity and reflation. Nobody has a playbook for the other thing. Yang, whatever else he got wrong, at least wrote the policy.
So the likeliest break is the least cinematic. The flows do not reverse the way the scenario needs, because they did not in 2009 and did not in 2020, and the burden sits on the argument to show why this time differs.
But something real sits underneath the numbers I have spent this essay discounting, and it does not depend on any of them. For two centuries we automated the work of people who owned nothing, and the argument about what we owed them was a moral one that markets were free to ignore. This time the displaced own the asset. That is true today, at current employment, before a single forecast comes good.
The retirement account was supposed to be the thing that made displacement survivable. It turns out to be the wire that carries it.