Abstract

Will new technology consign human workers to the garbage dump of history? Prior waves of concern over technological mass unemployment have typically come during periods of, well, mass unemployment. During the depths of the Great Depression, manufacturing automation became a target of policy makers, labor unions, and mass media condemnation. As one cartoon from that era presciently asked, “Is the robot beginning to think?” (Bix 2000). Amidst the halting recovery from the Great Recession, fears of robotization and technological substitution reappeared (Acemoglu and Restrepo 2020; Handel 2022). In the intervening 80 years, waves of unemployment and inequality had brought periodic attention to the way that ingenious new tools could, more or less temporarily and more or less locally, make our skills obsolete more quickly than they could generate new demand and new uses of human ingenuity. And if technological progress has never brought permanent mass unemployment to developed countries, it has delivered winners and losers and substantial adjustment costs (Autor 2015).

Our current economic context is not fertile ground for fears of technological unemployment. In the US, we are near historic highs in prime-age employment rates, workers are reaping the gains of a very long–running tight labor market, and workers in low-wage jobs have seen particularly strong wage growth. This heady mix has meant the first decline in wage and earnings inequality since 1980 (Aeppli & Wilmers, 2022). After decades of unequal growth, overall labor market inequality has not grown consistently since around 2012 and has declined markedly since 2015 (Dey et al. 2022). This marks the first concerted fall in labor market inequality since the 1970s, and since 2020, it has been at a scale that rivals the Great Compression of the 1940s (Autor et al. 2023). Indeed, this has been a period of particularly strong demand for precisely the non-college workers who bore the brunt of information technology’s labor-replacing and surveillance-increasing effects. The labor market is generating strong demand and rising real wages for workers at the bottom of the labor market.

And yet, notwithstanding this economic context, seemingly inhospitable to technological pessimism, the drama and apparent magic of generative artificial intelligence (AI), and large language models (LLMs) in particular, have launched a new wave of debate about the labor market effects of technological advance. Will this new technology mean a return to rising inequality, weak demand for workers, and intensified winner-take-all labor markets? Will it finally usher in the mass unemployment that has concerned technology commentators since the Great Depression?

In the following essay, I first summarize the scarce research we have so far on the likely labor market inequality effects of generative AI, with a particular focus on LLMs. I then bring lessons from broader research on employment and labor markets to outline potential channels through which generative AI could impact inequality. Throughout, I emphasize ways that standard pessimism about technology effects on inequality may be misplaced in predictions about new generative AI tools. Complex effects on the overall labor market and on inequality are difficult to predict but could be quite different from previous general-purpose technologies. Considering a rich set of channels—not just immediate labor substitution or skill complementarity but also changes in organizations and collective action patterns—reveals multiple ways that inequality could be affected by widespread generative AI adoption. I hope that attention to these multiple possible channels will increase the sophistication and scope of the emerging debate over generative AI’s effects on labor markets, workers, and inequality.

1. What Do We Know So Far?

A substantial area of research addresses effects of new technology on labor market inequality. However, this research, largely consolidating in the 1990s and 2000s, has focused on the rise of computers, information technology, the internet, and related technological advances. These studies find that these technologies substitute for routine tasks (Autor and Dorn 2013; Acemoglu et al. 2023b) and raise returns for highly skilled workers (Bound and Johnson 1992). Automating some production processes reduces demand for the assembly workers who previously worked on them, while introducing computers into the office raises the productivity of managers and professionals.

The construction of internet infrastructure offers a well-studied example of technological effects on inequality. Broadband internet roll-out improves labor market outcomes for high-skilled workers, while worsening outcomes for lower skill workers (Akerman et al. 2015). It exacerbates inequality between rich and poor localities (Forman et al. 2012). Likewise, firms implementing new information technologies increase demand for skilled labor (Bresnahan et al. 2002), and technologies like these facilitated the rise of superstar firms (Autor et al. 2020b). Different research designs, addressing different facets of information technology and considering related but distinct outcomes, all suggest that these technologies have tended to increase inequality by concentrating benefits on already advantaged workers, firms, and cities.

However, inequality effects of computers and the internet are not a clear guide to potential effects of generative AI. These are distinct technologies with potentially disparate effects. Moreover, the broader labor market context of prior studies of technology adoption is one that could exacerbate inequality effects: weak labor unions, import competition disproportionately affecting non-college workers, and stagnating minimum wages. Research on technology effects on labor supply and demand often takes for granted these contextual and institutional features, which could be important scope conditions for the inequality-increasing effects identified in the literature (Parolin 2021). In our current period of emboldened unions, rising employment regulation, and strong bargaining power for low-wage workers, these institutional-contextual effects could play a quite different role in the diffusion of generative AI.

So will generative AI have different effects than its well-studied information technology precursors? Unfortunately, direct research on labor market effects of generative AI is nascent and speculative. There is not yet enough of a track record on AI to make strong claims about its real-world inequality effects. However, two suggestive findings on inequality emerge from the research so far.

First, the tasks that generative AI, and specifically LLMs, replace are those typically done by workers in high-paid occupations. One research team categorized occupations based on the tasks that workers in those occupations typically perform and then rated tasks according to how substitutable with LLMs they are (Eloundou et al. 2023). This exercise requires some heroic assumptions: strong complementarities between tasks must be ignored; substantial within-occupation heterogeneity and task segregation could change the scale of job impacts. But, nonetheless, the results are striking: for low-wage occupations, those making around $30,000 per year, around 15 percent of their job involves tasks exposed to replacement by LLMs. For high-wage jobs, making around $100,000 per year, tasks accounting for nearly half of their work are exposed to potential replacement. Thus, it is the higher-paid, typically white-collar, occupations that are most exposed to generative AI. Another approach to predicting employment and earnings effects of generative AI also finds that it is likely to displace and reduce earnings for primarily (though not exclusively) white-collar workers (Kogan et al. 2023).

These preliminary results are not surprising: nonroutine cognitive work involving thinking, writing, planning, or coding tends to be high wage. And these are precisely the kinds of work tasks that LLMs are distinctively capable of performing and augmenting: ChatGPT is writing our emails, memos, business plans, and codebase. Compare these effects to industrial robots or self–check-out registers that chiefly automate the work of lower-wage workers. In those cases of automation and information technology, repetitive, manual tasks like painting a car or operating a cash register are replaced in part by automation.

Generative AI may thus have a very different skill-bias than prior innovations. It is white-collar positions, rather than blue-collar, service, and other frontline jobs, that are composed of tasks most immediately exposed to disruption by generative AI. If this disruption involves more task substitution than augmentation, then at the level of broad occupation categories, generative AI tools could reduce demand in white-collar jobs relative to blue-collar and in-person service jobs.

The second key finding from research on generative AI so far concerns comparisons of workers doing similar tasks and jobs, rather than comparisons of workers in quite different occupations. Within-job, access to LLMs tends to advantage the lowest-performing, rather than top-performing, workers. When customer service agents were given access to a GPT chatbot, they were able to resolve customer problems more quickly. But these efficiency benefits were only felt by the lowest-performing agents; top-performers experienced little improvement (Brynjolfsson et al. 2023). The top-performing customer service agents may have already implemented the effective tips and tricks proposed by their new AI assistant. Or they may use subtler response patterns, more tailored to particular problems and customers, than even the AI can propose. Regardless, the added tool did not push out the productivity frontier for top agents but rather helped close the gap between high and low performers.

This pattern is not limited to agents chatting with disgruntled customers. When law students were given access to ChatGPT for a legal case–analysis exam, the lowest performers on prior exams reaped the most benefit in improved scores (Choi and Schwarcz 2023). Similar results hold for professional consultants doing a variety of client service tasks: generative AI tools helped the consultants propose more compelling ideas for new shoe designs, effectively segment markets, and write persuasively about their proposals (Dell’Acqua et al. 2023). But this help provided much larger benefits for the consultants who were worse performers on tasks similar to these, substantially closing the performance gap between top- and bottom-half consultants. Finally, access to generative AI benefited freelancers asked to do professional writing tasks like press releases, emails, and analysis plans (Noy and Zhang 2023). But it also substantially compressed their productivity distribution, demonstrating another example of lifting the bottom toward the top. Across these different jobs and tasks, LLMs appear to be a skill leveler, delivering the biggest performance boosts to the previously worst performers.

Note that some of these equalizing effects could be regression to the mean. Low performers on an initial task tend to improve their rank on a subsequent task. Apparent equalizing in these early studies could also depend on early uptake effects: if no one is really great at using a new piece of technology, then maybe the most accessible, lowest-hanging benefits will accrue to those who have the easiest room to improve. LLMs may eventually, rather than leveling a playing field, simply create new dimensions across which skill can be unevenly cultivated.

Likewise, the research on generative AI exposure for white-collar workers cannot fully distinguish substitution from complementarity effects. If generative AI technologies raise exposed workers’ productivity more quickly than substituting for their work tasks, then this could deliver disproportionately faster wage growth in those occupations. However, for example, the study of freelancers cited above found that their use of ChatGPT was largely substitutive rather than complementary: freelancers just copied and pasted output without effectively adding value through editing and human input (Noy and Zhang 2023). Some preliminary evidence suggests that the mix of labor-saving and labor-augmenting effects will play out unevenly across specific occupations (Kogan et al. 2023), even among white-collar jobs. These results are suggestive that generative AI will have a substantial human labor substitution component for some white-collar jobs. But real ambiguity remains about the likely balance of substitution and augmentation across occupations.

However, in the following, I assume that some version of white-collar demand reduction and within-job leveling or compression effects will hold for future LLM implementation. If these early patterns persist, how will generative AI affect labor market inequality?

2. Technology Strikes Back

The simplest, first order effect implied by generative AI’s task ambit is that it could have a substantial leveling effect on labor market inequality. By replacing managerial and professional workers, it could diminish pay gaps based on skill. Indeed, the college wage premium has already stabilized since around 2000, after two decades of rapid growth (Valletta 2018). Following the slow recovery from the great recession (Aeppli and Wilmers 2022) and accelerating during the COVID-19 recovery (Autor et al. 2023), the college wage premium actually began to fall. Likewise, gaps between higher-paid and lower-paid occupations had begun to close by around 2015 (Aeppli and Wilmers 2022). These trends indicate some mix of the exhaustion of the skill-biased information technology revolution (Beaudry et al. 2016), tightening labor markets after the dislocations of the Great Recession (Aeppli and Wilmers 2022), and some shifts in employer practices and treatment of frontline jobs (Wartzman 2022). This is the context of generative AI’s deployment.

If generative AI further dampens relative demand for college-educated workers, because it substitutes for tasks central to managerial and professional jobs, this would lock in these declining inequality trends. Indeed, there could be a growing swath of tasks for which non-college workers, armed with generative AI tools, could effectively replace more expensive college-educated workers. Simple, clearly patterned diagnostic tasks could be shifted from doctors and nurses to medical assistants. External-facing business writing tasks could shift from managers and communications specialists to administrative assistants (Noy and Zhang 2023); more and more legal analysis could move from partners to associates to paralegals. These expertise-intensive jobs, often augmented by new information technology, have delivered decades of robust pay growth for managers and professionals, heightening gaps between white-collar occupations and the bottom half of the labor force. Generative AI may mean technology strikes back against its former beneficiaries.

Put broadly, if generative AI truly substitutes for, rather than simply complements, human intelligence, it challenges the economic dominance of the high-achieving, highly educated workers whose rising pay has defined US labor markets since the 1980s. The kind of problem-solving and pattern recognition that defines the traditional professions, but also broad swaths of managerial and software development work, could be taken on by highly effective generative AI.

This scenario requires continued rapid improvements in AI effectiveness and reliability. Indeed, there are many possible barriers to drastic change in these powerful jobs: technological, social, and institutional. But there was also a time of feudal order in which productive farmers paid tribute to armed gangsters. In our current system, that tribute is paid to elicit the human capital and know-how of white-collar workers. We incentivize the use of intelligence and expertise in the direction of increasing convenience and abundance. This deal so central to our labor market context—announced in 1990s rhetoric about the New Economy, the creative class, the information-based economy, and skill premiums—that it can seem natural and unassailable. But the ascendant professional and managerial class could face its most significant challenge yet in the form of generative AI.

3. The Erosion of Meritocracy

Beyond these between-occupation dynamics, generative AI could challenge the fine-grained rankings and distinctions that support highly unequal earnings outcomes for workers doing similar tasks. Insofar as generative AI accelerates the diffusion of effective writing, thinking, and pattern recognition from top performers to laggards, and corrupts standard performance assessment tools, it could undermine the claims to urgent meritocracy that justify much inequality.

Consider the competitive selection processes that undergird a range of institutions in the contemporary economy. Schools and elite universities, professional services firms and technology companies, are sorting machines, in which a key goal is sorting the very best applicants and performers from the herd of middling aspirants (Domina et al. 2017). Information technology contributed to, and was shaped by, this rise of meritocracy. Shifts from standardized pay to performance-based pay required improved worker monitoring (Massenkoff and Wilmers 2023; Pierce et al. 2015; Aral et al. 2012). More broadly, superstar pay-off dynamics for firms, teams, and individuals have all been linked in part to the implementation of new, skill-biased technologies (Autor et al. 2020b; Song et al. 2019). In the context of large differences in pay-offs between seemingly similar workers in the same occupation, or firms in the same industry, careful performance assessment becomes worth investing in.

But many of these selection processes have focused on assessment not through intelligence quotient and personality tests, as early commentators predicted (Young 2017), but rather relied on open-ended assessment that allowed cultural capital to shine (Bourdieu 1984). Interviews can cover not only technical know-how but also the hobbies, dispositions, and fit that mark elite candidates (Rivera 2012). College essays must reveal the right balance of dogged perseverance and fragile creativity (Karabel 2005). Yet, these verbal assessment processes are precisely those which generative AI coaching can already effectively simulate and hack. For all the concern about algorithmic bias, generative AI could level the playing field on the kind of soft cultural exclusion devices that are rife in elite boundary policing.

Yet, even more deeply, the meritocratic regime is only cost-effective if the productivity differences between slightly different candidates are large. If generative AI continues to be an effective skill leveler, then it could disrupt this hyper-selective ethos in a more thoroughgoing way than it challenges epiphenomenal meritocratic assessment processes. Generative AI could accelerate the diffusion of the habits, insights, and approaches of top performers to everyone. They could help low-productivity firms adopt the strategies and analytics of high-productivity firms. They could provide new, adaptive channels for learning and skill development. If these changes accumulated across a broad range of tasks, generative AI would begin to undermine the underlying basis for hyper-meritocratic selection processes.

4. Collective Action Possibilities

These labor market and organizational incentive changes could be complemented by shifting political coalitions. A long line of research in political science emphasizes how exposure to employment insecurity can support preferences for social insurance (Estevez-Abe et al. 2001). Workers exposed to employment risks are more likely to demand insurance and redistribution against economic losses. Consistent with this research, exposure to technological job loss among white-collar workers may strengthen a political coalition to protect workers from employment instability.

A comparison to another employment shock is telling. When job losses due to international trade were concentrated on non-college production workers, these effects were minimized and ignored for decades. There is little evidence that legislators’ responses to trade conditions actually affected legislation passed (Feigenbaum and Hall 2015). Under these conditions of political inefficacy, trade- and immigration-related grievances eventually exploded in a polarized and populist politics (Autor et al. 2020a). The transmission of shifted political preferences into effective policy short-circuited.

However, this slow and destructive political burn may reflect a general lack of political responsiveness to lower-income voters (Bartels 2016). If advances in generative AI throw higher-income software engineers, project managers, and doctors out of work, this could provide a basis for a more politically influential coalition for employment protections or redistribution than the prior period of negative pressures on already lower-income workers. Indeed, in a political system tilted toward higher-income voters, exposing some of these voters to economic risk may bolster inequality-reducing redistribution.

Beyond electoral politics, generative AI could drive a new wave of workplace collective action. Labor unions thrive when there is bargaining for collective goods in the workplace (Wright 1987). Absent union representation, workers have an incentive to free-ride on these collective goods and push instead for individual benefits from which there are no positive externalities. As such, the demand for unions can be strengthened by urgent demand for workplace collective goods among workers. Craft unions protected work rules and defended barriers to entry that benefited all incumbent members. Industrial unions emerged governing health and pension benefits. Bargaining over wages and salaries alone can rarely maintain a coalition: higher-productivity workers are tempted to cut self-interested side-deals, while lower-productivity workers who benefit the most bring the least bargaining clout to a union. Workplace collective goods provide a basis for concerted union support.

Governing generative AI use and deployment could be a key collective good across many occupations. Rules restricting worker replacement by AI would benefit a large swath of workers in threatened occupations. But, the benefit of these rules is too dispersed for them to emerge from individual worker bargaining. As with other workplace collective goods, unions offer a possible vehicle for collective action toward governing generative AI use. The Hollywood strikes of 2023 provide examples. An actors strike led by the Screen Actors Guild won language restricting the use of actors’ likenesses without their permission (as in AI-enabled digital effects). Likewise, an agreement following the Writers Guild of America strike stipulated that production companies could not require writers to use AI tools and discourages replacing writers with AI (Coyle 2023).

If these early examples persist and spread, revitalized labor unions could heighten wage compression in the labor market (Western and Rosenfeld 2011). At the same time that unions protect collective goods in the workplace, they compress wages among members. As such, even if it is higher-paid, white-collar unionization that is primarily stimulated by generative AI, this could strengthen inequality-reducing effects.

Of course, collective action is complex and difficult to predict. Nonetheless, these possibilities for new political coalitions and for revitalized unions offer another channel through which generative AI could affect inequality. Notwithstanding the speculative nature of the link between generative AI and complex collective action issues, these non-market and non-organizational channels can have meaningful second order effects on labor market inequality.

5. Reasons for Pessimism

The channels discussed so far all suggest that the ways that generative AI differs from previous general-purpose technologies (and specifically information technology) may drive continued reductions in labor market inequality. However, there are several channels through which even an effectively white-collar–substituting, low-skill–augmenting generative AI could nonetheless drive a return to rising inequality.

Most simply, the labor markets for college and non-college workers are interconnected. During the Great Recession, there were substantial occupational downgrading and skill mismatch. College graduates took jobs as baristas, lawyers as administrators (Lu et al. 2022). Declining demand for managers and professionals can mean increased labor supply competing for frontline jobs (Beaudry et al. 2016). Insofar as higher-education workers can outcompete lower-education workers for non–white-collar jobs, declining demand for managers and professionals could push down wages for blue-collar jobs. Absent new job creation to absorb these excess workers, technological unemployment could reverberate across many types of workers.

Beyond these pure labor supply effects, generative AI may reshape the underlying distribution of high- and low-paying firms. One recent article measured predictive AI adoption using AI experience on resumes at publicly traded companies (Babina et al. 2021). Companies that hired many AI experts were larger and in more concentrated product markets. These companies experienced more growth via product innovation. These results suggest that initial AI adoption could strengthen winner-take-all firm-driven inequality. However, this study did not address generative AI specifically and focused on AI development expertise rather than on the wide adoption of tools like chatbots. We can only guess at the long-term effects of generative AI on inequality between high- and low-paying firms.

Finally, generative AI could have inequality-increasing effects on workplace monitoring and employment relations. In a masterful ethnography of investment bankers, Karen Ho found that their experience of employment insecurity and risk led bankers to impose similar conditions on their portfolio companies (Ho 2009). Above, I noted how the experience of insecurity and expendability could shift white-collar workers’ political preferences. But, the process Ho identified is also plausible and perhaps more immediate. If the managers, investment analysts, and consultants who govern the employment relationship for frontline workers come to see precarity as normal and natural, this could warp their treatment of subordinates. Rather than viewing the employment relationship as a long-term basis for common purpose, it would make the low road of low wages, high turnover, and heavy monitoring more plausible.

Perhaps most concerningly, generative AI could facilitate easier collection and processing of data on individual productivity and performance (Acemoglu et al. 2023a). One study of AI use in call centers found that AI-enabled cameras and perpetual real-time coaching intensified worker monitoring (Doellgast et al. 2023). This kind of surveillance system can undermine efficiency wages and bargaining power for low-wage workers. Managerial implementation thus becomes critical for assessing workplace-level inequality effects of generative AI adoption: a race between skill leveling and declining costs of surveillance could determine whether generative AI undermines or locks in intensive monitoring and selection systems.

6. Predicting the Future

These multiple possible channels do not allow reliable predictions about the total effect of generative AI on labor market inequality. Even simple economic models of task substitution and labor demand require more data than we currently have on these novel technologies. Going beyond those simple models, to consider collective action, skill leveling, and workplace surveillance, takes us far beyond the realm of reliable prediction. However, working through these multiple pathways should warn against simple assumptions about the likely inequality effects of generative AI.

This approach can also guide policymakers to broaden the consideration set of ways that LLMs can affect inequality. Even for narrow, domain-specific technologies, like programmable machine tools (Noble 1978), priorities in technology development and the details of implementation shape their effects on workers. Apocalyptic visions of mass unemployment make for compelling media coverage. But they obscure many more realistic and potentially more malleable channels, through which generative AI could affect labor markets and the workplace. If generative AI could shrink gaps in pay between white-collar and production or service workers, or dull the hard edge of winner-take-all meritocracy, it could actually help lock in a more egalitarian labor market.