A new economics paper argues that competition will push companies to automate faster than is good for any of them, and that only state intervention can stop it.
Chief executives of leading artificial intelligence labs routinely predict that the technology will replace most human jobs. A new academic paper asks a question those forecasts tend to skip: who will buy the goods if workers are no longer paid?
In “The AI Layoff Trap”, Gerry Tsoukalas and Brett Falk build a mathematical model of that problem. Its central finding is uncomfortable. Even chief executives who see the danger and want to avoid it may be unable to, because the logic of competition pushes each of them to automate regardless.
The mechanics of the trap
Prof Tsoukalas, who discussed the research in a recent interview, says the 60-page paper can be reduced to a simple thought experiment. Imagine a single monopolist making widgets, with its own workers as its main customers. Offered a technology that would let it dismiss 90 per cent of its staff, it would hesitate, since those staff buy its products. In this case, he says, the model shows the firm “does just enough” automation and stops short of destroying its own demand.
Competition changes the calculation. With many widget makers, each firm’s workers spread their spending across the whole market. A company that sacks staff loses only a small fraction of its own sales, while capturing the full saving in labour costs. A firm that holds back, meanwhile, risks being undercut by rivals that do not and going bankrupt anyway.
Economists call this a dominant strategy: automating is each firm’s best move whatever its rivals do. The result is a prisoner’s dilemma. If every company reasons the same way, aggregate demand erodes and all end up worse off. Prof Tsoukalas stresses that this does not depend on chief executives being foolish. “It’s not that they don’t understand the economics,” the interviewer put to him, and the professor’s answer was that the individual cost-benefit sum simply points the same way for everyone.
Why voluntary restraint fails
The authors liken the position to Odysseus, who had himself tied to the mast to survive the sirens’ song. Companies, they say, need to slow the replacement of people with machines. But the model predicts they will not do so on their own.
Nor, Prof Tsoukalas argues, would a gentlemen’s agreement work. If the leading AI developers and big technology groups agreed in a room to hold back, each would face an immediate incentive to defect on leaving it, knowing the others were restrained. “That’s your opportunity and that’s your duty to your shareholders,” he said. Any effective deterrent would need “sticks and carrots”, and the sticks “have to be pretty significant”.
What would work
The authors assessed about six policy ideas now circulating among economists, including universal basic income and giving workers equity stakes in their employers. According to Prof Tsoukalas, only one tackles the root problem: the incentive to swap a worker for a machine. His proposal is a tax on full replacement of staff by AI. Augmenting workers with the technology would stay cheap, but eliminating their jobs would carry a cost.
He concedes the idea has few precedents. The nearest is carbon pricing, which does not forbid pollution but makes it costly and lets firms offset it against government targets. Because “tax” is politically toxic, he suggests alternatives such as subsidies for companies that retain workers. These might be “not perfect”, he says, “but we could use” them.
A bleak counterfactual
Pressed on what the United States would look like in a decade without action, Prof Tsoukalas was cautious, noting that “no one agrees” on the question and that economists are split. His own view is that the technology will keep improving, which he says he is “very confident” about, and that the likely result is wealth inequality on a scale not seen before. That, he warned, could bring political instability, as the 2008 financial crisis did.
Caveats
Several points are worth keeping in mind. The findings come from a theoretical model, not empirical evidence, and the interview does not address how quickly AI could actually displace labour, or whether new jobs would emerge to absorb displaced workers. The model also treats demand as the main brake on automation, whereas critics may point to other forces, such as new products, falling prices or state spending, that could cushion the shock. Prof Tsoukalas himself acknowledges that opinion on the economic effects of AI is deeply divided.
Even so, the paper adds a sharper argument to the debate. If it is right, the usual reassurance, that markets will self-correct because bosses need customers, is precisely what cannot be relied on. “Waiting for the firms to figure it out for themselves,” Prof Tsoukalas said, “is the worst possible thing we can do.”