Standing in front of a Paris audience this week, Jeff Bezos offered the sunniest forecast yet about artificial intelligence and work. AI will not replace human workers, he argued. It will create a labour shortage. People have a near-endless list of things they want to build and make, AI lowers the cost of doing all of them, and the result is more work for humans, not less. Asked about the widespread fear that AI will hollow out the job market, Bezos said he "totally disagrees."
A lot of people do not share his confidence. A Reuters/Ipsos poll released the same week found that roughly half of American workers believe AI threatens jobs in their own household. So the two views now sit side by side in plain sight: the world's optimist-in-chief promising a hunt for more workers, and millions of those workers bracing to lose the jobs they already have. One of them is going to be wrong, and the honest answer is that we cannot yet tell which.
Bezos is leaning on a pattern that has held for two centuries. Electricity, the assembly line, and the personal computer all destroyed whole categories of work and still ended with more people employed than before, often in better-paid jobs. It is the backbone of every optimistic case for AI. But that pattern carries a quiet condition: the displaced workers, or their children, have to be able to move into the new jobs the technology creates. A textile worker in 1920 did not retrain as a software engineer. Their grandchildren went to college and did. The transition was real, and it took the better part of a lifetime.
That is the part the cheerful version skips. If AI meaningfully cuts into coding, customer service, knowledge work, and creative fields within roughly a decade, as many forecasters expect, there is no time for a generational handover. There is barely time for a single worker to retrain once. New categories of work will surely appear, in AI infrastructure, safety, supervision, and an ageing world's demand for care. But a coder displaced by an AI assistant cannot easily reinvent themselves as an AI safety researcher, and when the supply of willing workers outstrips the new openings, wages drift down rather than up.
What makes the argument so hard to settle is that the data is still ambiguous, and both camps can claim it. Surveys show AI adoption surging, with around four in five organisations now using it somewhere in the business, yet documented job losses remain tiny: one widely cited tally found fewer than 17,000 US jobs lost to AI over seventeen months. Optimists read that as proof the panic is overblown. Pessimists read it as the calm described in the old line about going bankrupt "gradually, then suddenly," and warn that a recession could turn slow adoption into a sudden rush to automate.
Tellingly, Washington has decided it does not actually know. On 17 June the Foundation for American Innovation announced an effort to sharpen the federal surveys that track AI's effect on jobs, aiming to measure displacement and creation with far more precision than the current guesswork allows. Canada and the United Kingdom have launched similar studies. It is an unglamorous admission, but a useful one: before anyone can plan retraining, safety nets, or Bezos's hypothetical shortage, they need to know what is really happening. Until those numbers arrive, the loudest voices in the debate are working from conviction, not evidence.