For Black Workers, the AI Risk Runs Through Education, Not Race

A truck driver. A customer service rep fielding calls through a headset. A nursing assistant, a cashier, a janitor pushing a cart down a corridor at closing time. These are among the ten occupations that employ the largest numbers of Black workers in the United States, more than 4 million of them all told, and a new Urban Institute brief lines each one up against a single number: how exposed the job is to artificial intelligence. The picture that falls out is not the one the headlines about AI and jobs usually paint.

Black workers, it turns out, are overrepresented in the jobs AI is least likely to touch. Not the most. That is the first surprise, and it complicates a worry that has been circulating since generative AI arrived: that automation will fall hardest, and first, on the workers already furthest from economic security.

The brief, by William Congdon, LesLeigh Ford, Claire Cusella and Fernando Hernandez-Lepe, is not a survey of what AI has actually done to anyone. It’s an analysis of exposure, which is subtler. The team took an exposure index built by Tyna Eloundou and colleagues, which scores each occupation from 0 to 1 by how many of its tasks an AI tool could plausibly speed up, and laid it over demographic data on roughly 150 million workers from the 2024 American Community Survey. It’s an observational analysis of that survey data, a grey-literature policy brief rather than a peer-reviewed paper. Exposure here means only that a task could be done faster with AI, not that anyone has lost a job, or will. The authors are careful, almost insistent, on this point: exposure is not displacement. A task getting automated is not the same as a role getting automated.

That distinction is baked into the exposure index itself. In the original Science paper, Eloundou and colleagues estimated that around 80 percent of the US workforce could have at least a tenth of their tasks affected by large language models, and about 19 percent could see at least half, while stressing that a task being technically exposed says nothing about whether it will be automated, augmented, or left alone. The Urban brief inherits both the measure and that caution.

Sort every occupation into three bands by exposure and the pattern is plain enough. In the lowest-exposure third, Black workers make up about 13% of the workforce; in the highest-exposure third, 9.6%. White workers run the other way.

The Exposure Follows the Schooling

Why would that be? The brief’s answer is where the story turns from race to something adjacent. Exposure, across occupations, rises with education. The jobs an AI can help with tend to be the ones that ask for a college degree: the mean wage climbs from about $40,800 in the least-exposed band to nearly $88,600 in the most-exposed, and the college-educated share climbs right alongside it. Black workers, meanwhile, are overrepresented in occupations with lower levels of formal education, a pattern the authors tie to long-documented occupational segregation and crowding. Put those two facts together and the lower exposure looks less like a lucky break and more like an old inequality wearing new clothes.

Run a plain regression of AI exposure on the share of Black workers in an occupation and you get a negative slope, though one the authors call only marginally significant. Add controls for gender and, crucially, education, and the relationship all but evaporates: the coefficient on Black representation drops to roughly -0.036, statistically indistinguishable from zero. The link between how Black an occupation is and how exposed it is, in other words, is mostly the link between how Black it is and how much schooling it typically demands. Race is not doing the work in the data. Education is.

That result reframes everything around it, and is the point the brief’s own web summary glides past. If exposure tracked race directly, you would look for remedies aimed at race. Because it tracks education, and because education itself is shaped by decades of unequal access and occupational crowding, the risk sits one layer down, in how workers got sorted into occupations in the first place. And exposure, remember, cuts two ways. In a job with high exposure and high education, AI is more likely to augment what a worker does, handing them a faster tool. In a job with high exposure and low education, it is likelier to substitute for the work outright. The brief uses the college-educated share of an occupation as a rough stand-in for which way things will break. Sort the highly exposed jobs by that proxy and Black workers turn up disproportionately in the ones on the automation-prone side of the line. Customer service representatives are the clean example: heavily exposed at 0.567, 18.5% Black, and only about a third college-educated. Registered nurses, also highly exposed, sit on the other side, 94% college-educated, and there the same exposure reads as a tool rather than a threat.

None of this is a forecast. The exposure measure is prospective and uncertain, one of several such measures that broadly agree but never perfectly, and it says nothing about whether or when firms will actually adopt anything. Only time, the authors\ write, will tell whether exposure turns into real economic impact.

There is a second gap. Because they had reliable income data and no comparably reliable wealth data, the authors define the Black middle class by earnings alone, even while arguing that wealth is the better measure of who is actually secure. Roughly one in four Black households held zero or negative wealth in 2021. A class line drawn by paycheck misses that, and the brief knows it. The wider wealth picture only underscores the point: by Pew’s analysis of Federal Reserve data, the median Black household held about $27,100 in 2021 against roughly $250,400 for the median white household, and Pew separately confirms that 24 percent of Black households had no wealth or were in debt, the highest share of any group measured. Earnings capture little of that.

A Map, Not a Forecast

So what is the finding good for, if not prediction? It’s a map of where to look, and where to act, before the disruption arrives rather than after. The policy suggestions that follow are unglamorous and familiar: enforce fair hiring and investigate displacement that breaks along racial lines, shore up an unemployment insurance system that has never served Black workers well, and pour real money into retraining and apprenticeships, an area where the authors note the US ranks second to last among 38 wealthy countries in spending. That ranking is not an outlier claim: OECD figures put US spending on active labour market policies below 0.1 percent of GDP, in the bottom cluster alongside Chile, Iceland and Mexico, against an OECD average around 0.4 percent and well over 1 percent in the highest-spending countries. The goal, the authors are at pains to say, is not to keep Black workers away from AI. It is the opposite: to get more of them into the exposed-but-augmented jobs where the technology pays off.

Frequently Asked Questions