Factories Return but Do the Jobs

Capacity and employment are separate promises

A factory returning to the United States is evidence of domestic productive capacity. It is not, by itself, evidence that workers have recovered the employment, wages or bargaining power associated with an earlier industrial economy. The same distinction applies when a government restricts employers’ access to foreign workers: changing the recruitment channel does not determine how firms reorganize production.

The central claim is conditional. Domestic output can rise while labor required per unit falls. Whether total employment grows then depends on demand, new tasks, domestic sourcing and the pace of displacement. A successful industrial policy can improve resilience or strategic capacity even when its employment effect is modest. Those benefits should be stated separately rather than converted into an unsupported jobs promise.

This question belongs primarily in Volume IV, The Last Human Workforce, with connections to Volume II’s institutional power and migration questions and Volume III’s industrial restructuring. It examines policy and ownership; the companion workforce article explains task bundles and transition design.

The employer has more than two alternatives

An employer facing more costly or uncertain recruitment can raise wages, train domestic workers, redesign jobs, automate some tasks, move work overseas, reduce output or abandon an investment. These responses can coexist. The relevant question is which margin actually changes, not whether each restricted foreign position has a guaranteed American replacement.

Glennon’s research on historical H-1B restrictions finds that multinational firms increased employment in foreign affiliates, particularly where firms already had international operations. This establishes offshoring as an empirically supported response in that setting. It does not show that all firms offshore, that every rejected visa becomes a foreign job, or that permanent labor-certification restrictions have identical effects. [1]

The social-media claim that a green-card processing intervention means imminent mass deportation also needs a separate evidentiary chain. The Department of Labor describes PERM as labor certification generally obtained before an employment-based immigration petition. H-1B is a different employment category. A certification delay cannot by itself establish an individual’s status, a removal order, or an AI-related motive. Policy scope and individual immigration consequences must be verified independently. [2]

Compare the cost of reliable work

Would an American replacement cost more? There is no universal answer. Compare workers doing comparable tasks, at comparable skill and responsibility levels, in the same labor market. Include pay, benefits, recruitment, training, immigration administration, turnover, management and the cost of failures. Citizenship alone does not identify the cost difference.

The H-1B wage rule requires the higher of the prevailing wage or the employer’s actual wage for similarly employed workers. That legal requirement does not establish perfect compliance, but it does rule out treating lower wages as a universal feature of the category. A claim of underpayment needs employer-specific evidence. Nationality is also a poor proxy for a task’s technical exposure to automation. [3]

Compare automation on the same basis: cost per acceptable completed task, including integration, data preparation, security, review, maintenance and errors. A system that drafts slides or code cheaply may still require expensive domain judgment. Illustrative calculations are useful, but they are not payroll evidence: if output rises 20% and labor hours per unit fall 30%, total hours become 1.2 × 0.7 = 0.84 of baseline, a 16% decline. This assumes unchanged task mix, no supplier effects and no additional new work; it is arithmetic, not a forecast.

What the evidence can establish

Acemoglu and Restrepo’s task framework distinguishes displacement of labor from reinstatement through new tasks. Productivity can expand demand, while new human tasks can raise labor demand. Consequently, an argument about automation must account for both forces. Counting automatable tasks without modeling demand and newly created work cannot settle the employment question. [4]

The final 2025 publication of Generative AI at Work reports a 15% average increase in issues resolved per hour among 5,172 customer-support agents using an AI assistant, with heterogeneous effects. This is a concrete example of assistance improving productivity in one workplace. It is not evidence that those jobs disappeared or that every engineering role produces the same gain. [5]

Evidence also changes with tools and task selection. METR’s February 2026 update says its later developer experiment gives an unreliable estimate of current productivity because participation, submitted tasks and time measurement were affected by selection. Its earlier slowdown finding should therefore remain dated to its particular early-2025 setting rather than serve as a timeless verdict on AI coding. [6]

METR’s May 2026 survey of 349 technical workers reports median self-assessed value gains of 1.4–2 times, depending on the question. It explicitly distinguishes perceived speed from value and warns that counterfactual self-reports can overstate gains. Neither a benchmark, a worker’s impression nor a single workplace study is an audited estimate of economy-wide job replacement. [7]

Who captures the gains

The ownership argument is stronger when it names observable institutions. Who controls equipment, models, distribution, data and customer contracts? Who can change prices or work arrangements? Does compensation grow alongside productivity? These questions turn a broad claim about oligarchs into propositions that can be investigated through ownership, concentration, bargaining arrangements and the allocation of value added.

Autor and colleagues find evidence consistent with rising sales concentration in firms with lower labor shares contributing to the aggregate labor-share decline. This is a reason to track which firms capture growth, not proof that every large firm exploits workers or controls government. Political influence is a separate claim requiring evidence about lobbying, access, rules or enforcement. [8]

The argument does not require an enormous AI boom. Acemoglu’s 2024 model links aggregate gains to the affected share of tasks and task-level savings and presents modest, assumption-dependent projections. It also considers distribution between capital and labor. Its forecasts are historical analytical estimates, not measurements of October 2026. Even modest gains can raise a distributional question if ownership and bargaining determine who benefits. [9]

A test that can prove the argument wrong

The strongest objection is that automation may make a domestic plant viable when the alternative is no domestic plant. Higher productivity can lower prices, expand demand, preserve complementary jobs and support new suppliers. Recruitment enforcement may also address real violations regardless of employment effects. These possibilities require a counterfactual: compare the policy with a credible alternative, not only with a labor-intensive factory from decades ago.

Test the promise at both firm and plant level. Compare affected employers with similar unaffected employers before and after separately dated interventions. Measure operating employment rather than construction announcements, hours, real compensation, output, domestic value added, worker retention, training, contractor work, foreign-affiliate employment and documented task changes. Track subsidies and strategic capacity separately. Pre-existing layoffs, demand shocks, tariffs, interest rates and mergers can confound the comparison.

Support for the argument would be rising output accompanied by lower domestic hours, documented task substitution or expanded overseas work after restricted recruitment. Sustained domestic employment and real wage gains exceeding a credible comparison group would weaken the claim that the policy chiefly relocates assets rather than improves workers’ position. Poor pre-policy comparability limits causal conclusions; these outcomes alone cannot reveal officials’ or executives’ motives.

My policy proposal is to make public support conditional on transparent, multi-year reporting and enforceable worker benefits where feasible: training, redeployment, compensation and community returns. Report labor compensation relative to value added alongside total pay and headcount; a falling share can coexist with rising wages. The test of bringing industry home is what productive capacity the country gains, what durable opportunities people gain, and how the gains are distributed.