For years we were told that AI would do the dull work, spare us the drudgery and free humanity for higher things like poetry, social care and better coffee. Instead, one of the more plausible outcomes is that firms automate away large numbers of workers, wreck a slice of the consumer demand on which they themselves depend and then stare at the shrinking economy blaming everyone else.
That, in essence, is the argument of a striking new paper, The AI Layoff Trap, by Brett Hemenway Falk and Gerry Tsoukalas. It says rational firms, looking clearly at the rocky road ahead, can still put their foot on the accelerator.
That is what makes the paper so good. It does not depend on executives being venal, drug-addled, untrustworthy, fanatical or stupid, though one would not want to rule out the current field evidence. It depends on something more durable than madness and ketamine: incentives.
The point is simple. When a firm replaces workers with AI, it pockets the saving itself. That is the private benefit. But when those workers lose wages, they also lose spending power. They buy less of everything, from trainers to takeaways to streaming subscriptions. That lost spending does not just hit the firm that sacked them. It spreads across the economy. So each company gets the private benefit of automation while carrying only part of the social cost.
That social cost matters because workers are not only workers. They are also customers. If enough firms replace enough people quickly enough, they do not merely cut labour costs. They start to eat their own demand base. The paper’s point is that this can happen even when everyone can see it happening.
Put crudely, the AI revolution may be building an economy in which firms proudly replace their customers.
Picture the PowerPoint to Sam Altman. Headcount down. Margins up. Efficiency enhanced. Then, somewhere around slide 38, a minor deterioration in the continued existence of people with wages.
The paper goes further. It argues that more competition can make this worse, not better. A monopolist at least swallows more of the damage it causes. In a crowded market, each firm is more tempted to grab the saving and let the demand fall splash around the rest of the sector. That is how you get an automation arms race among people who think of themselves as hard-headed realists.
I’m sending this paper, with a whiff of alarm, to the Chancellor, in the hope it reaches the pointy heads in the Treasury and Peter Kyle. The paper deserves attention well beyond economics departments because it identifies the problem at the point of decision, not merely in the wreckage afterwards. It warns that the market contains no reliable brake.
In the paper’s frictionless case, the logic hardens into a Prisoner’s Dilemma. Every firm automates. Every firm ends up worse off than under collective restraint. No firm can escape by behaving nobly on its own.
I can see Mark Zuckerberg now, in the wood-panelled boardroom of his missile-proof super yacht, squinting into a Zoom with Elon Musk calling in from his space base, both men nodding gravely as they agree to automate the final solvent customer out of the economy. It is a marvellous image for our age, a virtual room full of very clever men congratulating themselves on their strategic brilliance while quietly setting fire to themselves.
That is why the paper is politically important. Too much of the AI debate treats the problem as something that happens later. People lose work, then government tidies up. Falk and Tsoukalas say the problem starts earlier. The real issue is not only what happens after displacement. It is the competitive incentive to cause the displacement in the first place.
That is awkward for nearly everyone.
It is awkward for the AI evangelists who talk as if labour were somebody else’s problem. The laissez faire right still hopes the market will sort itself out with more innovation.
For the pro-growth centre-left, this is a nasty complication. Competition, the usual corrective, is part of the problem here. The broader left is in no better shape. Our instinct is often to clear up after the market has done the damage: tax profits, strengthen labour, retrain workers, spread ownership and top up incomes. The paper’s argument is tougher than that. Those measures may soften the blow, but they do not remove the incentive to automate too far, because each firm still pockets the saving and pushes part of the wider demand damage onto everyone else. The left, in other words, cannot rely on a bigger ambulance alone. It also needs brakes.
The paper tests a whole shelf of remedies. Wage adjustment does not solve it. Free entry does not solve it. Upskilling narrows the problem but does not remove it. UBI does not solve it. Capital taxes do not solve it. Worker equity participation does not solve it. Coasian bargaining does not solve it. The authors’ conclusion is that only a “Pigouvian automation tax” can fully correct the distortion.
In plain English it means this: if companies do not naturally pay the full cost of the damage they impose on everyone else, government puts a price on that damage so the private calculation matches the real one. We do it, in principle, with pollution. The factory owner may enjoy the profit from dumping muck into the river, but the public pays for the filth downstream. A Pigouvian tax says, fine, you can do the thing, but you must also pay for the harm you were previously offloading onto the rest of us. Pollution tax.
Applied here, the idea is simple enough to explain. If a firm automates a job and pockets the saving, but the lost wages reduce demand across the economy, then some tax should claw back that unpriced damage. Not to ban useful technology, but to stop firms acting as if the only relevant line in the ledger is their own. The human version is simpler: if you insist on firing your customers, do not expect the rest of us to subsidise the experiment.
There is a darker cultural joke in all this. We have spent years sending bright young people to university so they can learn how to write crisply, present neatly, code competently, summarise fast and absorb institutional wisdom. Then one morning a board decides that Anthropic’s Claude can automate every polished quarterly task they used to go to university for, and suddenly the ladder into professional life is hauled up with an email about agility.
This does not make the paper anti-AI. Many firms are building tools that will improve public services, accelerate research and unlock scientific and medical advances that would have seemed miraculous a generation ago. Used well, AI will help us run the state better, discover faster and solve problems that have resisted human effort for years. The warning in the paper is not against innovation. It is against a market logic that can take a transformative technology and drive it towards a destructive outcome.
A technology that saves labour by deleting labour income may look brilliant on the spreadsheet and ruinous in society. Capitalism, in a fit of locally rational enthusiasm, may use the machines to wound itself.
It is enough to bring Len McCluskey back from retirement just to say, “I told you so.” Trotskyists, after all, have been predicting the fall of capitalism since, well, shortly after Trotsky stopped being available for comment.
And that is the really modern touch. We can model the trap, publish the paper, nod gravely at the findings and still march towards the cliff in excellent order, congratulating ourselves on our agility.
The question is not whether AI can help humanity do extraordinary things. It plainly can. The question is whether politics can shape the incentives before the market optimises itself into folly.
PS I did take a serious look at this dystopian possibility several years ago!