The digital divide stopped being a connectivity problem about a decade ago. It is now a labor-market problem — and treating it like a broadband-access issue is why federal policy keeps missing the target. The Americans who lack the tools to compete in an automated economy are not, for the most part, people who cannot get an internet signal. They are people who have a smartphone, a patchy plan, no home workstation, no employer-sponsored training budget, and a job that the U.S. Bureau of Labor Statistics' 2024–2034 Employment Projections classifies as exposed to either automation or task-level AI substitution. The fix is workforce policy, not fiber.
That distinction matters because the trillion-dollar investments unlocked by the Infrastructure Investment and Jobs Act and the CHIPS and Science Act are now largely deployed, and the broadband piece is finally moving. According to the National Telecommunications and Information Administration, the $42.5 billion Broadband Equity, Access, and Deployment (BEAD) program had approved final proposals for all 50 states and U.S. territories by mid-2024. Yet Pew Research Center's 2024 surveys still find that 7% of U.S. adults do not use the internet, concentrated heavily among Americans over 65, those earning under $30,000, and those without a high school diploma. Those groups overlap almost perfectly with the workers most exposed to displacement.
The divide is now about complementary capital, not connectivity
MIT economist Daron Acemoglu has spent the last several years arguing that the productivity gains from automation flow to workers who own complementary capital — the skills, devices, and discretionary time to direct machines rather than be directed by them. In his 2023 NBER working paper with Simon Johnson on the political economy of AI, he frames this as the central question of the decade: not whether machines will do more work, but who gets paid for the work that remains. The empirical picture is consistent with that frame. McKinsey Global Institute's 2023 "The Economic Potential of Generative AI" report estimated that 60–70% of employee time in current occupations involves activities that could be automated by 2030 under aggressive adoption — but the same report found the productivity uplift accrues overwhelmingly to workers in the top quartile of digital fluency.
What that means on the ground: a logistics dispatcher with a dual monitor, an employer-paid Microsoft Copilot license, and twenty hours of training is now roughly twice as productive as a peer doing the same nominal job on a personal laptop with no AI tooling. The wage gap between those two workers, in our reading of BLS Current Population Survey tabulations, has widened sharply since 2022. The divide isn't internet access. It's whether your employer treats you as worth augmenting.
Three policy moves that would actually close the gap
If the goal is to make sure no worker is left behind in the automated economy, the interventions that show the strongest empirical track record are not the ones policy debate keeps gravitating toward. They are, in rough order of evidence:
1. Portable training accounts tied to the worker, not the job
Singapore's SkillsFuture Credit program — which gives every citizen over 25 a personal training wallet usable across thousands of accredited providers — has been studied extensively by the OECD and is the closest thing the world has to a working model. The 2023 OECD Skills Outlook called it "the most replicable individual learning account scheme currently operating at national scale." A U.S. equivalent, sized at $2,000 per worker per year, would cost roughly $300 billion over a decade. That is less than the IRA's clean-energy tax credits and would reach four times as many workers.
2. Employer co-investment requirements for AI deployment
The empirical literature on firm-level AI adoption, summarized in Erik Brynjolfsson and colleagues' 2023 work on generative AI in customer-support settings (published in the Quarterly Journal of Economics), shows that productivity gains accrue when workers receive structured training alongside the tool. When firms deploy AI without training, the gains collapse and worker turnover rises. A regulatory floor — say, two paid training hours per AI license deployed — would internalize that externality without requiring new bureaucracy.
3. Apprenticeship rebuild in the trades AI cannot do
Roughly half of the jobs the BLS projects to add the most positions between 2024 and 2034 — home health aides, nurse practitioners, electricians, wind-turbine technicians, solar installers — are physical, regulated, and largely AI-immune. The U.S. has about 640,000 registered apprentices today, versus more than 1.5 million in Germany on a population a quarter the size. Closing that gap is the most concrete thing federal workforce policy could do this decade. NWLB's flagship work on this is in Apprenticeship 2.0 →.
What the data says we should stop doing
Two interventions absorb a disproportionate share of workforce-development funding and have weak evidence behind them. Short, online-only "digital literacy" courses with no employer pathway show negligible labor-market returns in the OECD's PIAAC longitudinal data. And one-off retraining grants disbursed without wraparound supports — childcare, transportation, income replacement — have a completion rate below 30% in most state-administered Workforce Innovation and Opportunity Act programs, according to the Department of Labor's own evaluations. The Brookings Institution's 2024 review of WIOA outcomes concluded that the program's per-participant earnings impact is "statistically indistinguishable from zero" without those supports.
That is not an argument for less workforce spending. It is an argument for spending it on programs the evidence says move people into better jobs.
The digital divide is no longer about who has internet. It is about who has an employer that treats them as worth augmenting — and who has a government willing to fund the training when their employer won't.
The next five years will decide which model the United States adopts: the Singaporean approach of investing in the worker directly, the European approach of regulating how firms deploy AI, or the current American default of letting outcomes sort themselves out by ZIP code. The first two work. The third is what produced the divide in the first place.
Updated May 21, 2026. This piece was substantively rewritten as part of NWLB's 2026 editorial refresh.
