Most of what I write here begins with a running curiosity about what's happening in capital markets, which means I read a lot of forecasts!

Last week, the New York Times published a story on a question I keep circling in this newsletter. What is AI doing to the economy, and how would we know? By some measures, the technology is feeding graduate unemployment and has already destroyed tens of thousands of jobs. By others, the companies using it most are hiring faster than the cautious ones. Researchers can't even agree on how many firms are using AI in the first place.

And yet the big forecasts keep coming. Goldman Sachs estimates that 300 million jobs worldwide are exposed to AI automation and that 6-7% of American workers could be displaced over the next decade. These numbers are shaping headlines, academic research, and now legislation in Congress, and nearly all of them depend on the same underlying idea, a metric economists call AI exposure.

Roughly, "exposure" is an estimate of how much of a job's tasks today's AI could plausibly perform. One of the main reasons economists constructed these measures is that official government statistics are slow to update. The most detailed occupational data the government publishes comes out just once a year. The latest release only arrived this May, pushed back by last fall's government shutdown, and it describes the job market as it stood in May 2025, fourteen months ago. That's a long lag for a technology moving this fast. Building a proxy to fill the gap was a reasonable thing to do, but the issue now is that the proxy is carrying a lot of weight, arguably too much, in current discussions.

Start with the fact that the measures don't agree with each other. Back in 2013, researchers Frey and Osborne asked ML experts to judge whole occupations and concluded that 47% of US employment was at high risk of computerization. When the OECD re-scored the same question task by task, the number fell to 9%. The newer measures built around large language models haven't converged either. The Yale Budget Lab recently compared seven of the leading metrics and found that they agree about who is "safe," with plumbers sitting near the bottom of every list, but diverge at the top.

But even a perfect exposure score wouldn't tell you what happens to wages or employment. David Autor has a comparison I think about a lot.

In his "Expertise" model, an air traffic controller and a crossing guard do roughly the same job, keeping vehicles from colliding with people, yet the controller earns about 4x as much. The difference is that controller expertise takes years to build, while almost anyone can fill in as a crossing guard. That scarcity is what the market pays for, and it's also what determines what AI does to a given job. If AI automates the abundant, low-expertise work, the people holding rare skills become more valuable. If it automates the rare, high-expertise work, the barriers to entry fall, competition floods in, and wages come down. Two jobs can have identical exposure scores and completely opposite futures.

On top of that, the disruption these scores are supposed to predict hasn't shown up in the aggregate data yet. Once again, the Yale Budget Lab went looking for a connection between AI exposure and actual labor market outcomes, and once again, as of its June update, came up empty. There's no clear AI footprint in the occupational mix and no link between AI usage and unemployment. The sharpest finding on the other side, the Canaries paper, showed a 16% relative employment decline for the youngest workers in the most exposed jobs, but with caveats from its own authors about timing and about what else was happening in the economy. This point I've covered here before; things could dramatically change, but we haven't seen that yet.

There's also a difference between what AI could do and what it's being used to do. The gap between the two is shaped by adoption, cost, and how long it takes firms to reorganize around a new tool. Anthropic's economic research team found a version of this in their own data. The theoretical exposure score on its own showed no correlation with projected job growth by occupation, while a version weighted by how people use AI in practice did. What could happen turns out to be a weak guide to what does.

All of this is why I'd rather see us put less weight on any single measurement. Policymakers, naturally, want to offer constituents some certainty amid the anxiety, and legislation is starting to follow. The AI DATA Act, which Senator Mark Kelly introduced with two Republican colleagues in June, would modernize the federal labor surveys and require an annual report on AI's effects on the workforce. Whatever the merits of that particular bill, the instinct to measure better is at least pointed at the real problem. Writing a contested score into policy triggers and eligibility rules would not be, especially while the measures still disagree this much.

It's simply too early to know which future we're getting, the one where AI mostly substitutes for workers or the one where it mostly augments them, the fast one or the slow one. And if AI really is a general-purpose technology, the second- and third-order consequences will be the ones no exposure score can see. Electricity's biggest effects showed up in factories, cities, and workdays rebuilt around it, decades after the lightbulb, and nobody scored any of that in advance. We aren't going to diagnose our way to the right future ahead of time. The better investment is anticipatory capacity, systems and strategies built to hold up across several futures at once, so that when this technology surprises us, and it will, we can adapt rather than scramble.

If you want to dig into these questions further, Bharat Chandar, one of the co-authors of the Canaries research, keeps a running list of what we know and don't know about AI and labor markets. Check it out!

Our partners and sponsors get exclusive placements across the newsletter and access to AIC's in-person network — demo nights, dinners, hackathons, and forums across 180+ chapters.

For all inquiries, send us a note at [email protected].

The AI Collective is built by volunteers across 180+ chapters in 40 countries.

Thank you to the thousands of volunteers around the world who make this work possible. We truly could not do this without you.