Yale Budget Lab: No Labor Market Disruption from AI Yet

The Yale Budget Lab has been tracking the impact of AI on the labor market since early 2026, publishing monthly updates based on Current Population Survey data and usage metrics from OpenAI and Anthropic. Their April 16 update, covering data through March 2026, delivers a clear verdict: no measurable disruption yet.

The lab measures three things. Occupational dissimilarity, which tracks how much the mix of jobs changes month to month. Industry dissimilarity, the same measure at the sector level. And AI exposure and usage metrics, drawn from OpenAI’s task-exposure estimates and Anthropic’s actual Claude usage data. Across all three, the picture is stability.

“Occupational dissimilarity, industry dissimilarity, and our exposure and usage metrics all remain flat, lie within historical ranges, or continue along the trends they were already exhibiting,” the lab writes. The single exception is a small uptick in dissimilarity between younger and older college graduates, but that remains at the high end of the historical range, not outside it.

This is worth taking seriously. The Yale Budget Lab is not a think tank with a thesis about AI. It is a policy research group at a major university, publishing raw data and letting the numbers speak. The numbers say that five years after ChatGPT’s launch, the occupational mix is changing at roughly the same pace as it was in 2021, before any generative AI product existed.

OpenAI’s exposure data, which estimates what share of tasks in each occupation could theoretically be performed by generative AI, shows no shift in employment shares across exposure quintiles since ChatGPT launched. Workers in the highest-exposure occupations are not shrinking as a share of total employment. Workers in the lowest-exposure occupations are not growing.

Anthropic’s usage data tells a similar story. The lab uses Anthropic’s February 2026 usage metrics, based on a sample of one million conversations from Claude and the enterprise API. The proportion of employment in occupations with high levels of task-level AI usage, whether automation or augmentation, has been stable. Even among unemployed workers, the share of tasks that could be automated shows no clear difference by duration of unemployment. If AI were driving layoffs, the recently unemployed would show higher exposure scores. They do not.

The lab’s framing is careful. They note that these are early data points and that the effects of transformative technologies often take years to show up in aggregate statistics. “It is too soon to tell how disruptive the technology will be to jobs,” they write. “The lack of widespread impacts at this early stage is not unlike the pace of change with previous periods of technological disruption.”

That last sentence is doing real work. Every general-purpose technology in history, from electricity to the internet, took a decade or more to reshape labor markets. The personal computer revolution of the 1980s did not show up in productivity statistics until the mid-1990s, a lag that Robert Solow famously called out in his 1987 quip: “You can see the computer age everywhere but in the productivity statistics.” AI may follow the same pattern.

But the Yale data also raises a harder question for the AI industry. The dominant narrative from frontier labs and venture capital is that AI will automate knowledge work at scale, displacing millions of white-collar workers within a few years. That narrative is not supported by the data through March 2026. If the disruption is coming, it has not started yet.

The Anthropic usage data is particularly telling. The lab reports that observed usage is more likely to be associated with automation than augmentation. That means when people use Claude, they are more often having it complete a task entirely rather than using it to enhance their own work. Yet even that pattern of substitution has not produced a measurable shift in employment. The tasks being automated appear to be small enough, or distributed enough across occupations, that they do not change the aggregate numbers.

One interpretation is that AI is automating tasks, not jobs. A worker who uses Claude to draft emails or summarize documents is still employed. The task is automated, but the role remains. That is consistent with what the Yale data shows. It is also consistent with what economists call the task-based model of labor markets, where technology changes the composition of tasks within occupations rather than eliminating entire occupations.

The lab’s analysis has limitations. They note small sample sizes for some subpopulations, particularly recent college graduates. The Anthropic data covers only Claude users, not the broader AI tool landscape. And the exposure metrics from OpenAI are theoretical, not actual. But the overall pattern is robust across multiple data sources and measurement approaches.

For AI builders, the Yale Budget Lab’s findings suggest a mismatch between product ambition and market reality. The industry is building systems designed to replace workers. The labor market is treating those systems as productivity tools that leave employment largely unchanged. That gap may close over time, as models improve and deployment scales. Or it may persist, if the actual economics of automation turn out to be more complex than the narrative suggests.

The lab plans to continue updating this analysis monthly. That is the right approach. The question is not whether AI will affect labor markets. It almost certainly will. The question is when, how, and at what scale. The Yale data through March 2026 says the answer is not yet.