Calls for governments to push “pro-worker AI” sound appealing. The idea is simple: If policymakers deftly guide how the technology develops, they can make sure it helps workers instead of replacing them. What’s not to like?

Here’s your trouble: Technology almost never works that neatly. Its effects on jobs are usually messy, unpredictable, and shaped by millions of decisions from businesses and entrepreneurs—not by a policy plan designed in Washington.

That’s a core point in a recent critique by economist Joshua Gans of a proposal from Daron Acemoglu, David Autor, and Simon Johnson to steer AI toward worker-friendly uses. Gans says the idea runs into a basic contradiction. The proposal defines “pro-worker” technology as something that makes human capabilities and expertise more valuable. But those things are valuable partly because not everyone has them. If a new technology spreads skills more widely, it may help more workers overall—while at the same time reducing the pay advantage of those who once had rare skills.

Gans illustrates the tension with education itself. Universal schooling expanded literacy and improved skills, empowering millions of workers. But by reducing the scarcity of those skills, it also eroded the wage premium once earned by the few who possessed them. By the paper’s logic, even education would have “mixed” effects—an implication that underscores how unstable the definition becomes.

History backs up that skepticism. Trying to design technology around specific job goals can backfire. It risks locking in today’s assumptions about work instead of leaving room for entirely new kinds of jobs to emerge—the kind that often appear only after a technology spreads.

Many inventions that looked like job-killers at first ended up creating new kinds of work. ATMs were supposed to wipe out bank tellers, but instead they made branches cheaper to run, so banks opened more of them and shifted tellers into customer-service roles. Spreadsheets didn’t eliminate accountants. They created new careers in financial analysis. And GPS didn’t just replace map-reading, it helped launch ride-hailing, delivery apps, and modern logistics systems.

If you go further back, railroads were supposed to replace canal boats and the people who worked on them. And they did. But they also created entirely new kinds of jobs—from managing large companies to running nationwide shipping networks.

(A Hayekian point seems relevant here: Government planners lack the specialized knowledge and real-time market information to both pick winning technologies and understand all their longer-run impacts. Private actors, closer to the ground, are better placed to spot them.)

Before policymakers try to guide AI innovation in certain directions, their economist adviser should ask a basic question: What’s the market failure that needs fixing? With pollution, for example, the problem was clear—companies weren’t paying for the environmental damage they caused. But with AI, as Gans explains, it’s much less obvious what’s being mispriced or distorted. Without that identification, it’s hard to know why government should be picking which AI technologies to push or discourage.

That doesn’t mean government has no role. It can help workers adapt by investing in education and training, keeping markets competitive, and strengthening safety nets. It’s also worth remembering that if AI really is a powerful general-purpose technology, it will boost worker productivity. And higher wages will eventually follow—the most pro-worker impact of all.