The public debate around the expected labor impact of artificial intelligence (AI) has settled into a familiar pattern: headlines warn that AI is poised to eliminate millions of jobs, destabilize labor markets, and recreate the dislocation of past technological revolutions. But the emerging empirical record tells a different and more nuanced story.
Early examinations of various job categories suggest that AI’s labor effects to date have been small in the aggregate, uneven across occupations, and shaped overwhelmingly by institutional context. How firms deploy AI tools, how workers are trained to use them, and how policymakers structure the environment around adoption appear to be the most important factors. That is to say, the technology itself is not dictating outcomes—institutions are.
This distinction matters, because the United States has already lived through labor-market shocks that popular narratives have typically blamed on technology or globalization, even as the evidence later demonstrated that institutional failures explained important portions of the damage.
The canonical example is the “China shock.” When import competition surged in the 2000s, many U.S. manufacturing centers experienced deep and lasting job loss. The devastation arose from the way institutions failed to adapt to the scale and speed of the trade shock, rather than from any intrinsic feature of trade flows. Workers did not relocate, because U.S. mobility had been declining for decades. Retraining programs were too weak to redirect workers into new occupations. Local safety nets were inadequate. As a result, regional labor markets adjusted remarkably slowly, with wages and labor-force participation remaining depressed for more than a decade after the shock.
Economic shocks can become enduring scars when institutions fail to evolve with technological or global change. And this history provides the correct analytic lens through which to view the arrival of artificial intelligence. Rather than the routine-task automation of the late 20th century, AI is instead a cognitive-augmentation technology that can arguably raise the performance of lower-ability or less-experienced workers more than it does experts. It could also expand the geography in which high-value work can occur, and reduce the cognitive and financial fixed costs of entrepreneurship. All these features make AI a potential equalizer, rather than a destabilizer—that is, if institutions allow the technology’s benefits to diffuse broadly.
What matters for policymakers is the institutional environment through which AI diffuses. If training systems stagnate, mobility remains constrained, or regulation locks in incumbent advantages, AI could replicate the scarring patterns of earlier disruptions. By contrast, there is an opportunity to build complementary institutions that allow a general-purpose technology to lift productive capacity across the economy.
AI’s labor impact is not destined to resemble the China shock. It will only resemble it if we repeat the same institutional failures.
Lessons from Past Shocks: It Was the Institutions, Not the Technology, that Doomed Workers
If any episode has shaped modern anxieties about economic disruption, it is the China shock. Yet the core lesson of that period is routinely misunderstood. What devastated hundreds of U.S. manufacturing towns in the 2000s was not technological inevitability or the abstract force of globalization. It was, instead, a failure of institutions to adapt to a large and sudden shift in the composition of traded goods.
David Autor, David Dorn, and Gordon Hanson document the extent of this institutional failure. In response to import competition, many affected labor markets did not reallocate workers into new industries or new regions. Instead, they find the adjustment was “remarkably slow,” with depressed wages, low labor-force participation, and elevated unemployment persisting for more than a decade. The underlying reason was not a lack of available jobs elsewhere, but a severe decline in geographic mobility and a lack of institutional support to help workers transition.
U.S. mobility had been falling since the 1980s, the very period in which workers needed regional flexibility to respond to globalization. Interstate migration and job switching declined steadily for decades, driven in part by labor-market rigidities, housing constraints, and overall weaker local dynamism. As mobility fell, the burden of adjustment shifted increasingly onto communities that lacked the tools to respond.
The consequences were predictable but not inevitable. Regions exposed to Chinese import competition saw spikes in disability claims, long-term joblessness, and social distress. These are patterns that mirror the geographic concentration of harm, including rising rates of labor-force-participation dropout, rather than a nationwide technological displacement. These pathologies were strongest where retraining systems were weak, safety nets were thin, and workers could neither move nor transition into new occupations.
The result is a clear causal chain: geographic immobility, combined with weak institutional support, leads to persistent labor-market scarring, even when national economic conditions are otherwise healthy. The China shock showed the durability of labor-market harm turns on the strength of surrounding institutions. Where those institutions faltered, workers were far less able to withstand rapid technological or global shifts.
Inequality in the IT Revolution Was Created by Policy, Not by Automation
A similar misconception clouds discussions of the information-technology revolution. A common narrative holds that the rise of computers and automation in the 1980s and 1990s generated “skill-biased technological change” (SBTC), contributing to widening wage inequality by rewarding high-skill workers, while displacing routine ones. But the empirical foundations of that narrative have eroded significantly.
One systematic review synthesized 127 empirical studies published between 1988 and 2021. The evidence demonstrates that the timing and magnitude of rising U.S. wage inequality does not align with the standard SBTC narrative. Thus, fears of widespread technological unemployment are unfounded, because “the labor replacing effect of technology is typically more than offset by a range of compensating mechanisms suggesting that the widespread anxiety over technology-driven unemployment lacks its empirical base.” Across nearly all technology categories, the net employment effect was neutral or positive, with 29% of studies showing positive net effects, versus only 18% with negative net effects, and the largest share were highly dependent on context.
Indeed, most of the technology-related job losses observed in decades past reflected task reallocation, not the elimination of occupations. Technologies tend to shift employment toward nonroutine, service, and higher-skill jobs, while also creating new complementary roles. This evidence directly contradicts the idea that automation in the 1980s through the 2010s produced large, economywide job displacement. Instead, as the authors emphasize, “there does not appear to be an empirical foundation for the fear of technology-driven unemployment.”
At the same time, technological change can have disproportionate effects across skill groups by harming some while benefiting others. This is more consistent with institutional failures to manage transitions than with technologically predetermined labor displacement. For example, Europe—with very similar technologies but stronger labor institutions—did not experience the same inequality spike as the United States.
What Makes AI Different: Cognitive Augmentation, Not Substitution
A striking feature of today’s AI systems—particularly large language models (LLMs)—is that their primary economic effect is not to replace workers, but to augment human cognitive performance. This is especially true for workers with lower baseline skills or less experience. This characteristic sharply distinguishes AI from the popular SBTC narrative of the computerization wave of the 1980s and 1990s. Early empirical research consistently finds that AI raises the floor of performance, compressing productivity differences within occupations and enabling less-experienced workers to perform at levels closer to seasoned professionals.
For example, in one study with college-educated professionals performing writing tasks with and without ChatGPT, access to the model reduced task-completion time by roughly 40%, and significantly improved output quality. But the distributional pattern is even more important: the largest gains accrued to low-ability workers, pulling the bottom of the performance distribution upward and reducing variance across workers.
Another study confirmed this pattern in a controlled experiment linking ability, self-assessment, and AI usage. It found that “AI improves performance more for people with low baseline ability,” and that the gains are largest for workers who correctly assess their own limitations (those “calibrated” to rely on AI where appropriate). In other words, AI acts as a skill-leveling tool: novices are lifted toward the mean.
Taken together, these findings invert the popular SBTC account that new technologies are assumed to complement only high-skill labor. With AI, the comparative advantage appears to lie not in preexisting expertise but in the ability to collaborate effectively with a model.
Real-World Evidence: AI as an Equalizer
Emerging field studies reinforce this equalizing dynamic across several professions. In a large-scale deployment of a GPT-powered assistant at a Fortune 500 firm, productivity increased 14% on average. The case study’s results concluded that “AI assistance disproportionately increases the performance of less skilled and less experienced workers across all productivity measures.” As the study’s authors note, the system effectively captured and redistributed the problem-solving patterns of top performers.
In a study of small and mid-sized accounting firms, AI tools for classification, reconciliation, and error checking reduced routine data-entry work by 8.5 percentage points; allowed accountants to devote more time to client communication and higher-value analyses; and shortened the monthly close process by 7.5 days. Output quality also increased, as reflected in more granular ledgers and fewer inconsistencies.
In randomized controlled trials with law students and junior legal practitioners, AI tools dramatically improved both speed and quality. Complex legal tasks like drafting briefs, analyzing complaints, or preparing persuasive letters were completed 38–115% faster, with measurable improvements in reasoning depth, when the students and practitioners used advanced AI-reasoning models. Importantly, AI operated as a cognitive amplifier, enhancing human judgment rather than replacing it.
Across these domains, the pattern is the same: AI expands the capabilities of less-experienced workers, narrows skill gaps, and diffuses high-quality expertise throughout an organization. Thus, the early evidence suggests that AI is structurally a floor-raiser: its largest and most reliable productivity gains accrue to the workers who historically benefited least from new technologies. Whether this equalizing potential is realized at scale, however, depends almost entirely on access, training, and diffusion. In other words, it will depend on institutional decisions, not technological inevitabilities.
AI and Remote Work May Reverse Small-Town Declines
One of the most overlooked labor-market shifts of the past five years is the geographic rebalancing triggered by remote work. For decades, high-wage knowledge work remained tightly clustered in large metropolitan areas, leaving rural and small-town labor markets structurally disadvantaged. That pattern finally broke during the COVID-19 pandemic-era expansion of remote work. Approximately 59% of rural counties that had been losing population began growing again between 2019 and 2021, a reversal directly correlated with the rise of remote employment opportunities.
AI works in harmony with this geographic decentralization by making remote workers more productive in exactly the kinds of tasks that dominate modern knowledge work: writing, editing, synthesizing information, coding, analysis, customer communication, and design. In an analysis of post-2020 adoption of organizational technology, firms that had built remote-work capacity were also the fastest adopters of generative AI. This created an “organizational technology ladder,” whereby remote work forced companies to digitize workflows, and those digitized workflows made AI-integration frictionless.
This complementarity cycle means that the very tasks most suitable for remote execution are also the most AI-augmentable. A worker in a small town can draft legal memos, produce software prototypes, generate marketing collateral, or manage customer pipelines with the same AI-accelerated tools used in major metropolitan firms. AI dramatically narrows the capability gap between workers with access to dense local professional networks and those outside them.
If remote work loosened the geographic tether, AI has the potential to sever it entirely. A rural accountant, paralegal, developer, or marketing specialist can now contribute to industries historically concentrated in major metropolitan centers without relocating. For regions that suffered during the trade and automation shocks of the 1980s–2000s because of geographically isolated workers, the combination of remote-work infrastructure and AI-based augmentation represents a rare structural tailwind.
Labor Outcomes Aren’t Written in Code, They’re Written in Institutions
The picture that emerges from contemporary experience with AI and historical experience with labor disruptions is surprisingly consistent. AI does not behave like the automation technologies that animated earlier fears of technological unemployment. It is not a force that widens skill gaps or concentrates productive capacity in a handful of superstar cities or firms. Instead, AI appears to raise the floor of human performance by lifting novices, extending the abilities of workers historically disadvantaged by geography or educational background, and lowering the cognitive and financial barriers to entrepreneurship. If diffusion is allowed, AI has the structural characteristics of an equalizer.
But this equalizing potential is not automatic. The risks that matter today are not driven by technology but by institutional failures. A collapse in entry-level hiring pipelines; hiring systems overwhelmed by low-quality signals; state-by-state regulatory fragmentation; licensing and credentialism that restrict who can work with AI; and compliance burdens that only incumbents can bear are all potential issues that could convert AI from a broadly shared productivity enhancer into a mechanism that entrenches the advantages of large firms and high-status workers.
Conversely, with the right institutional scaffolding focused on portable skills, competitive markets, open standards, broadband access, and a regulatory posture that encourages experimentation rather than freezing labor-market structure, AI can diffuse across the entire economy. It can extend high-value work into smaller firms and smaller towns. It can provide new mobility pathways for workers who missed out on earlier technology waves. And it can allow entrepreneurship, not incumbency, to define the trajectory of the next decade of productivity growth.
The lesson from past shocks is clear: institutions, not technology, shape the trajectory of economic outcomes. When training systems, mobility supports, and regulatory frameworks fail to adjust, shocks harden into scars. When they adapt, general-purpose technologies generate broad opportunities. AI’s impact on workers will follow the same logic.
