The figures were discreetly delivered, tucked away in a Goldman Sachs report that most people ignored after skimming it. AI is cited as the main cause of the monthly elimination of about 16,000 jobs in the United States. Not a forecast. Not a caution. a current approximation. However, if you listen to Jeff Bezos speak at a technology conference in Paris, the picture appears to be completely different: there is an impending labor shortage, productivity is rising, and there are more jobs than people to fill them.
Neither of them can be entirely correct. And Stanford’s new Labor Lab is setting its flag in precisely that tension, the gap between those two narratives.
The lab, which focuses on modeling employment disruption before it shows up in national statistics, operates under a premise that seems almost out of style in the current AI optimism: the workforce warnings that are coming in now should be given serious, methodical attention rather than a wave of assurances from people with something to sell. Conversations with researchers in this field give the impression that the nation is being asked to have faith that things will turn out as they always have, and that no one has bothered to see if this time is actually different.
In the long run, Bezos might be correct. In the past, industrial revolutions have produced more jobs than they destroyed. However, that framing ignores a crucial aspect: the transitional phase. What became of the interim employees? How much time did it take? Who paid for it? In order to answer those questions, Stanford researchers are using actual data to monitor which job categories are declining the fastest, which areas are quietly absorbing the shock, and which demographics are slipping through the cracks before any official alarm goes off.
One early indicator is recent college graduates. Since 2022, that group’s unemployment rate has increased to almost 6%, growing twice as quickly as the overall workforce. The traditional entry point into professional careers, entry-level white-collar jobs, are disappearing in ways that are not readily apparent in headline unemployment statistics, which are still close to historic lows at about 4%. The total appears to be in order. The picture that has been broken down reveals a different narrative.
The number of tech layoffs through May 2026 has already exceeded 115,000, getting close to the total for the entire year of 2025. AI-driven efficiency has been used by Meta, Amazon, and Snap to justify cuts. According to a survey of chief financial officers, job losses due to AI could be nine times greater in 2026 than they were in 2025. These are not statistics from the periphery. They sit side by side with the more general optimism and do not reconcile; they just uncomfortably coexist.
The goals of Stanford’s lab are less ideological than they might seem. It’s not saying that automation should slow down or that AI is bad. The objective is more akin to what epidemiologists do: create a surveillance system that can identify an outbreak before it surpasses the capacity of the response. research that is anticipatory rather than reactive. There is currently no clear answer to the question of whether institutions are willing to act on early signals, which will determine whether or not that model succeeds.
No one else appears to be doing this at scale with the appropriate rigor, which is what makes the project truly fascinating—possibly even essential. There are forecasts everywhere. It is less common to cross-reference actual hiring data with actual modeling of displacement timelines, regional impact, and sector-by-sector erosion. It’s still unclear if the lab’s predictions will be accurate enough to be implemented or if there will be the political will to do so.
Building a significant early-warning system for the labor market doesn’t seem like a bet against progress at a time when half of American workers tell pollsters they fear AI will cost them or someone close to them a job, and when Anthropic’s CEO has called the impending disruption “unusually painful.” It appears to be the bare minimum of a sensible precaution.
