Labor is at the center, both real and imagined, of discourses and material processes of this thing we call AI. The topic is, however, difficult to research and write about because large gaps remain between narratives, promises, and reality. The problem of hype, in part, stems from the blurring of audiences with disparate expectations, for example, the board public, business management, the financial sector, and policymakers.
Much of what gets discussed as AI is just class conflict by another name.Footnote 1 A major strain of the marketing campaign around AI suggests that employers can stop hiring human beings entirely, in favor of more docile machines. Meanwhile, workers frequently complain that AI rollouts are intensifying and speeding up labor rather than replacing it.Footnote 2 That workers feel helpless to negotiate their pay, hours, and working conditions has nothing to do with AI per se. That is downstream of the weakness of workers’ organizations, compared to the strength of employers, and the larger regulatory environment, which makes employer speedups, outsourcing, classification changes, and splitting up work tasks to create “unskilled” labor easy to achieve.
A strange Potemkin AI incident recently involved a company that specializes in consulting for landlords who faced complicated situations. This company sold a service to these landlords that was billed as an “AI agent.”Footnote 3 This company, in fact, hired PhDs to pretend to be an AI that did the accounting and consulting for the landlords. The PhDs were compensated fairly well for their labor, but the service was marketed as labor-free. Many of the marketing professionals selling enterprise AI have come to the same conclusions about the preferences of their customers for nonhuman labor. Karen Hao’s recent book on AI reflects this trend—she describes AI companies looking for highly educated workers to supplement, shape, or simply rewrite AI outputs.Footnote 4 Of course, this arrangement is predicated on the existence of pools of underemployed but highly educated labor; a product of larger trends like the contraction of well-paying jobs in higher education. Journalists at 404 Media similarly reported the widespread use of decently paid ($50/hour), human contractors who read user prompts and do a lot of work improving “AI” output, including fact-checking.Footnote 5
None of this is to say that AI-based automation is a hoax. Certain work functions that had not been automated before may now be—especially in coding, legal document generation (e.g., contracts, wills), and other office work that follows standard formats and patterns. The question of how to socially reproduce this complex intellectual labor, if the need for it declines to a great extent, remains to be answered: for example, if mid-career lawyers remain necessary to oversee the construction of and evaluate legal documents, where will mid-career lawyers come from once entry-level positions at law firms are eliminated?
Much more is at stake, however, in debates about AI and labor than the elimination of labor tasks or jobs. One frequent concern about the present monomaniacal focus of investment on AI is that it crowds out other developmental paths by monopolizing physical materials, labor, land, time, and attention. Massive sums of money—government spending, retirement funds, industrial research expenditure, non-profit investments, and private business investment—are currently reorienting themselves around AI. What would they have done otherwise? What are the implications of not focusing our resources on fighting exploding inequality, climate change, or anything else? Might understanding the broader forces at work contribute to labor struggles, as well as wider social struggles, to redirect labor and flows of resources in emancipatory directions?
Toward a labor history of AI
This special feature of ILWCH comprises six articles. Four reveal both continuities and innovations with semiconductor industry labor regimes in the US, Europe, and Asia. (Silicon Valley has moved to even more global labor pools in the present for reasons that will become clear.) The final two articles add to histories of digital human labor.
The focus on the semiconductor capital history points to its centrality in plowing the grounds—ideological, material (e.g., supply chains), legal (e.g., trade regimes), and institutional (e.g., industrial policy and state support, industrial labor practices)—that has enabled the “AI” industry as it exists today. The semiconductor industry was singularly important in the drama that unfolded in the 1980s and 1990s between two visions of government—the new right and a diffuse coalition involving groups like Atari Democrats who thought government had a central role in high-tech led development. At the time of Reagan’s election, the Cold War security state and scientific government attached to it needed to be reconstructed. Correctly or not, political, academic, and industry elites saw the semiconductor industry as essential to the future of both US civilian economic and military interests.
In particular, the semiconductor industry responded to declines in market share in the early 1980s by vigorously lobbying Washington to do something about Japanese competition. In doing so, they produced a significant and unprecedentedly hostile policy response to an ally, Japan (e.g., an entirely new trade policy).Footnote 6 More importantly, this panic provided a stage and leverage for the technophilic coalition to score several victories over the New Right through the Reagan administration.Footnote 7 This group successfully built new institutions and transformed existing ones in line with their new theory of government. The trials of the semiconductor industry dominated politics—especially the debate over industrial policy—for the next two presidencies. The Clinton administration embodied the technological constituency perfectly and came into office with grand visions of state transformation. Again, the semiconductor industry played a massive role: a very public and much-lauded return of us companies’ prominence (especially Intel) ratified the success of this developmental model, and encouraged its spread to other technology and non-technological fields. This success, despite the victory of New Right avatar Newt Gingrich in 1995, largely also legitimized the theory of governance lobbied for by the technology coalition and stabilized the new high-tech corporatist military Keynesian state.
An important part of the semiconductor model was its approach to labor—one that administrations from Reagan to Clinton materially supported.Footnote 8 The industry experimented with forms of unfree, semi-free, and hypermobile labor arrangements.Footnote 9 They were among the first to experiment with offshoring, along with the textile industry. Our four contributions to this history come in two pairs—the first pair examines uses of unfree labor and the second pair examines union activity. In both pairs, one focuses more on labor in the 1970s while the other looks at more recent developments.
Ella Coon in “Progressive Corrections: Digital Electronics, Carceral Labor, and Liberal Reform in the 1970s” and Davide Blotta in his piece “From Indenture to Electronics: Brokered Labour Regimes in Johor, Malaysia” detail the ways in which semiconductor labor regimes build upon much older traditions of unfree or semi-free labor. Coon’s piece adds to a long literature on the use of the 13th Amendment’s infamous loophole embodied in prison labor.Footnote 10 This was a bipartisan effort that began, Coon shows, in the 1970s. Private use of prison labor had been banned during the New Deal. As Coon details, in 1973, Minnesota was the first state to allow private industry to pay prisons to use their facilities as a way to deal with spiraling state and municipal budgets. The federal government began legalizing private prison labor in 1979. Control Data Corporation (CDC), an important semiconductor firm through the 1980s located in Minnesota, was centrally involved in this transformation. They saw in prison populations an opportunity: cheap labor. This involved classifications of prisoners in terms of their fitness for electronics production and an argument about electronics as “light industry” well-adapted to prison conditions. The progressive vision touted by proponents was that “training” prison labor in the industries of the future would help them reform and eventually integrate back into society. When prison populations exploded through the 1980s and 1990s, William Morris, CDC’s CEO, brought this vision into federal policymaking, shaping present prison labor practices.
Blotta follows the legacy of colonial indenture to Malaysia, where he examines the migrant labor pools involved in chip packaging. Chip packaging is now largely done in Southeast Asia, though Biden’s CHIPS Act sought to reshore some of this capacity. This area is particularly interesting as one of the more labor-intensive parts of the semiconductor manufacturing process in the present day. Capital keeps those labor costs lower, Blotta argues, through modern iterations of older colonial indenture systems. These arrangements use “debt, document confiscation, sponsorship-dependent residence, and dormitory control” to bind workers. One particularly interesting point raised in this piece is that social reproduction in AI production may not be exclusive to knowledge labor and may create stumbling blocks for manufacturing capital.
Later Silicon Valley firms, including contractors who supply, e.g., labeling workers to AI companies, built on the practices and infrastructures pioneered by the semiconductor industry in this regard. Exploitation of refugees or other laborer pools in extreme straits is well documented.Footnote 11 The continued use of prison labor is worth further exploration. Recent reporting demonstrates that AI companies are indeed using prison labor without compensation by framing their labor extraction as biometric data collection.Footnote 12 The problem of socially reproducing docile labor in these cases seems to require constant mobility and exploitation of zones of desperation. The climate crisis and present energy crisis may provide such companies with more sites rather than fewer, making the imperative to organize across borders more acute.
Even when not exploiting legal loopholes in the prevailing taboo on slave labor, the industry has famously opposed worker organizing. Joseph Parziale’s piece “Speedup or Closure: The Global Shop Floor of 1970s Silicon Valley” and Yoehan Oh’s piece “Who Can Stop Semiconductor Production? Capital Vintage and Worker Representation at Samsung Electronics in the age of AI chips, 2018–2025” deal with union struggles globally in the post-1970s crisis paradigm of high-tech-led development. Parziale’s article details how low-cost labor and productive flexibility made a global electronics production system attractive in the 1970s. He argues, however, that the stability of this arrangement depended upon capital’s ability to preserve managerial control. Where workers acquired autonomous power, and firms had credible alternatives, relocation or closure could become a weapon against them; where firms were dependent on existing capacity, workers could gain leverage; and where authoritarian or corporatist institutions successfully contained worker power, firms could remain while tolerating controlled forms of union representation.
Oh’s contribution follows a similar path in examining the relationship between automation, labor strategy, and worker representation through Samsung workers’ union, NSEU, and their 2024 strike. This includes a valuable comparative history of Samsung, LG, and SK Hynix. The author shows that automation followed “capital vintage”: that is, cutting-edge chip fabs (the most valuable lines) are increasingly automated while older, less profitable lines remain more labor-dependent. This vintage created an uneven distribution of potential structural power within the firm, separating the sites of greatest labor leverage from those of greatest strategic value.Footnote 13 Together, these contributions clarify how mobility has been at the core of the semiconductor industry’s relative success with labor. Disciplining workers through mobility, or by selection of sites where regressive regimes meant political institutions containing worker power, became common practice. Where capital could not move, worker efforts had more leverage. More recently, workers at all three major memory chip manufacturers—SK Hynix, Samsung, and Micron—have used extreme AI demand for memory chips to fight for profit-sharing agreements. SK Hynix and Samsung have conceded to labor demands while Micron (as of this writing) faces strikes at their facilities in Taiwan.Footnote 14
The two final contributions gesture toward interesting further lines of inquiry away from more traditional factory work toward forms of digital human labor. In a 2023 internal memo, Microsoft executive Brent Hect called web scraping for AI production and maintenance “the greatest theft of labor in human history.”Footnote 15 While the veracity of this claim is arguable (slavery?), this alone would make web scraping a worthy topic for this issue. But Sebastião Quelhas Freire’s article “The History and Labour of Web Scraping” examines the non-automatable labor of web scraping itself. Joining the work of Science and Technology Studies scholars like Finn Brunton in his book Spam, Freire’s piece expands scholarship on arms-race-style adversarial digital labor like the work of spam filters.Footnote 16 Web scraping requires human expertise because websites, anti-bot technologies, Application Programming Interfaces, legal restrictions, and scraping technologies themselves are continually changing and adapting to avoid bots, Distributed Denial of Service attacks, and other kinds of negative or costly interventions. It remains to be seen whether, despite recent AI successes in illegallyFootnote 17 hacking systems, human labor will remain essential in other adversarial forms of digital work (e.g., hacking and cybersecurity). Freire’s article suggests yes, even with automation. In these kinds of adversarial systems, constant adaptation and change seem to require human oversight and labor.
Likewise, Lorenzo de Lelli’s “From the Economic Man to the TaskBot: A Genealogy of Rational Agents in the Evolution of Cyborg Science” offers an account of how economic concepts of rationality and rational agents influenced the parceling out of labor tasks in the kinds of data work involved in producing AI. As Stephanie Dick has shown, Simon’s bounded rationality changed significantly when it ran up against the limits of computing in the 1950s.Footnote 18 The implementation of this abstract idea in material infrastructure changed both idea and infrastructure. Not only are acts of translation from minds and machines really acts of transmogrification, but incorrect models can produce new infrastructures. Joe Lemelin and Jacob Browning in “Neural Networks,” for example, show how incorrect models of neurons influenced the design of neural nets. These neural nets (vast simplification) are the building blocks of modern GenAI. Specifically, they are, contrary to the beliefs of their progenitors, not good at Simon-esque rational reasoning but instead are more statistical-physics-like stochastic classifiers.Footnote 19 Where concepts of bounded rationality do shape AI, de Lelli argues, is in the design and distribution of human labor tasks in the many human data work inputs to AI. It is also interesting to contemplate the ways in which an abstract idea of a rational actor may shape the iterations of chatbots that have come to dominate a significant slice of public encounters with AI despite those models operating according to an entirely different logic (and ontology).
Avenues for future research
This special feature makes no claim to cover the issue of labor and AI comprehensively. The various forms of digital labor (cleaning, annotation, labelers, maintainers, output optimizers) often taking place in coercive, semi-free labor contexts have received much recent attention from journalists and academics.Footnote 20 Work on the quotidian labor happening at AI companies, along with marketing, integration labor, and their historical predecessors, appears much less common, but would be valuable. The consumption of AI in labor or as a labor input is the focus of much current scholarship and controversy but is also not covered here.Footnote 21
A future research agenda on labor and AI might examine how conceptions of automation, or beliefs about who is responsible and how, have changed over time. Similarly, a potential future line of inquiry would examine labor that goes into other conditions of possibility (or inputs) of AI; for example, the labor that goes into making massive political and infrastructural changes necessary.
Likewise, some work is being done on new labor movements cropping up in response to some of these new labor regime innovations (for example, the “AI as African Intelligence” movement).Footnote 22 Placing these in historical context would be useful. What kinds of labor/laborers are we producing under this regime? What does that mean for politics? What kinds of political subjects are produced and what politics are possible? Our articles on semiconductor labor innovations and responses may point in this direction.
Not much has been written about what to do with the presently technologically unemployed. Computer science majors and tech workers (especially those at or near entry-level) do seem to be experiencing significant job loss.Footnote 23 Perhaps future work could examine what government action has and should look like, including the problems of rebuilding government capacity and finding jobs for educated graduates and tech workers. Labor does not currently have a consistent position on AI; what should a unified labor position look like?
Behind AI merchant and government approaches to knowledge work and AI seems to be a naïve theory of knowledge with many implications for knowledge labor and labor history. This model of oracular knowledge assumes that there are universal disembodied facts that in theory could be gathered in one place or in one intelligence (a market, an LLM, etc.) and accessed via queries.Footnote 24 The history and philosophy of science and engineering, by contrast, includes significant work on the role of things like embodied and tacit knowledge, social reproduction of knowledge, infrastructures, circulations, networks, processes, practices, and social lives that produce useful, (somewhat) stable and durable knowledge through time.Footnote 25 These comprise one huge set of questions that loom over automation discussions but have not been explored enough. The social reproduction of labor and knowledge structures under this new regime alone is a massive and incredibly urgent topic.
We hope the articles in this special feature will contribute to the development of more robust research and thinking about labor and AI.
Open access
