The unemployment rate is not measuring what is happening to community college graduates

A new Brookings paper argues that AI is eroding the labour-market value of community college credentials, not by throwing existing workers out of jobs, but by drying up the entry-level positions those credentials were built to unlock. If that argument holds, the most-cited reassurance about AI and work, that unemployment among AI-exposed workers has not risen, is measuring the wrong thing for the population that has the least room to absorb the shock.

The paper. Harry Holzer, a Georgetown labour economist and former chief economist at the U.S. Department of Labor, and Amy Feygin of the American Institutes for Research published the piece at Brookings on 8 June.¹ Their claim is structural. AI adoption is reducing employer demand for entry-level workers in exactly the roles community college programmes train people for. The credential still exists. The signal it sends, this person is ready for that first job, is being decoupled from the job itself, because the job is being absorbed into software.

Why this matters more than it sounds. Community colleges enrol around 10 million students a year in the United States and educate roughly 40% of all U.S. undergraduates.² Their students are disproportionately from lower-income families, disproportionately first-generation, disproportionately people of colour. For most of them, the credential is not a status object. It is the route into a labour market that otherwise has no entrance for them.

The reassurance that does not reassure. In March, Anthropic published a labour-market paper using its own Claude usage data to construct a measure of AI exposure by occupation. The headline finding: "no systematic increase in unemployment rates for workers in occupations more exposed to current AI capabilities."³ That sentence has done a lot of work in the discourse since. It has been used to argue that the labour-market anxiety around AI is, at minimum, premature.

I do not think the Anthropic finding is wrong. I think it is being asked to carry a weight it cannot carry. Unemployment statistics measure people who already have a foothold in the labour market and lose it. They do not measure people who never get one. If AI is hollowing out entry-level hiring, the harm shows up first as graduates who do not transition into the field they trained for. That is invisible in the unemployment rate. It is highly visible in a 23-year-old's life.

The mechanism is signalling, not displacement. This is the move in the Holzer and Feygin argument that I think is worth pausing on. The standard worry about AI and work imagines a worker doing a job and an AI tool replacing the worker. The community college worry is different. It is about the first job, the one that converts a credential into a working biography. If employers automate that first rung — the junior paralegal task, the entry-level coding task, the first-year customer-support task — the rung is still there in the org chart, but it does not hire human beings any more. The credential that signalled readiness for that rung loses meaning regardless of how good the training was.

The data on entry-level hiring is not encouraging. Entry-level job postings in the United States fell roughly 34% between 2019 and 2023, according to Lightcast data.⁴ That predates the current wave of AI deployment, which means AI is not the originating cause. It is, plausibly, an accelerant of a structural shift that was already underway. The Federal Reserve Bank of New York found in 2023 that recent college graduates were experiencing underemployment rates around 41%, with community college graduates matching into the labour market less well than four-year graduates.⁵ The squeeze on the first rung is not a future scenario. It is the present.

What the consensus view gets right, and what it misses. Recent policy commentary, including from the Federal Reserve, frames this as a question of institutional adaptation. Community colleges, the argument goes, can preserve credential value by updating curricula, embedding AI competencies, partnering more closely with employers. Holzer and Feygin themselves call for more work-based learning, apprenticeships, employer-tied curricula, co-op placements, as the central policy response.

The curriculum response is correct as far as it goes. It is also, I think, insufficient on its own terms, for a reason that the recommendation itself implies but does not foreground. Work-based learning requires employers who want to host learners. If employers are automating entry-level functions because the economics of doing so are compelling, the same incentive that hollows out the first job also hollows out the apprenticeship slot. You cannot solve a hiring problem with a curriculum reform if the binding constraint is that the employer no longer wants to hire.

Who decides this. I keep coming back to a question that the productivity framing of AI deployment makes it easy to skip. When a firm automates its entry-level intake, the decision is made by the firm. The cost is paid by the cohort of would-be entrants, who were not in the meeting. This is not a complaint about firms behaving badly. It is a description of how the costs and benefits of a deployment decision are distributed. The firm captures the efficiency. The cohort absorbs the closed door. The community college absorbs the reputational hit when its placement rates fall. The credential becomes the thing that was supposed to work and did not.

What I would watch. Three things. First, first-job attainment rates for community college completers by field, tracked over time. The aggregate unemployment rate will keep telling a calm story; this number will not. Second, employer participation in registered apprenticeship programmes in fields with high AI exposure. Holzer and Feygin's prescription stands or falls on whether employers show up. Third, the wage gap between community college credential-holders who do enter the field and those who do not. If the entry-level rung is genuinely thinning, the people who clear it should command a premium, and the people who do not should be visible in the underemployment data.

The Brookings paper does not quantify the effect. It calls for the research infrastructure to measure it. That is the right call. In the meantime, the question worth holding is not whether AI is causing mass unemployment. It is whether a generation of people who did what they were told, get the credential, do the training, enter the labour market, is going to find the door they were trained to walk through has been quietly bricked up while no one was looking at that particular wall.

Footnotes

¶