Abstract
Keywords
At the dawn of what seems like a new technological epoch, policymakers and regulators find themselves at a critical choice point. On the one hand, they must position their constituents or stakeholders to leverage the new technology, ensuring employers’ and workers’ unfettered access to it. They will want to clear any potential obstacles between these economic actors and the bountiful economic and social fruits they aspire to cultivate. Cynical observers might note that policymakers, as politically animalistic as they are human, seek approval, especially from the elites and really anyone benefiting economically under the status quo. Thus, they may well glean as much psychic benefit from having guided a nation, state, county, or city to technological prowess as they would from any of the trappings arising from it. If instead we give policymakers the benefit of the doubt, we might ascribe their technology-forward inclinations to their desire for widespread prosperity: Clearly, economic value can diffuse only if it is created in the first place.
On the other hand, even technology-chasing, techno-optimistic policymakers must recognize the harm created by economic competition—turbocharged by novel technologies—in the absence of policy guardrails. The Industrial Revolution made factories thirsty for power, a need at least momentarily quenched by fast-moving belts, shafts, and flywheels. While these were unguarded at first, at least they were elevated, albeit along rickety, railingless walkways. At ground level, workers collaborated with the table saws and joiners that fed on this power. These technologies augmented both their personal productivity and their employers’ profitability exponentially. However, all too often, fingers, arms, legs, and lives became unintended, uncompensated, fully expended inputs in the production process. Even those policymakers basking in the glory and tax receipts, once confronted by supposedly muckraking journalists or labor activists, therefore had to acknowledge a regulatory mandate to mitigate risk. Pleasing factory owners alone would not be enough; the economic and social returns to new technologies, not to mention the costs arising from their use, would have to be shared across a broad set of stakeholders. Thus was the birth of the fundamental tension workplace and technology regulators face, even today, as embodied in a recent missive from the US executive branch.
Since taking office, President Biden, Vice President Harris, and the entire Biden-Harris Administration have moved with urgency to harness the potential of artificial intelligence (AI) to spur innovation and advance opportunity, while also taking action to ensure workers share in these gains. (White House 2024, emphasis added)
The White House’s focus has clearly moved on from the maiming machines of an earlier era. By now, blades and gears are coated, covered, and tucked away; machine speeds are regulated; and those toiling high above factory floors have something to hold onto or even strap into. Existing technologies now come with guardrails, figurative if not literal. But, what about those technologies just now diffusing across shop floors, offices, call centers, retail outlets, home offices, and any other conceivable workplace—not table saws or even smartphones, but AI and the algorithms, not to mention the fiber and 5G that connects workplaces to offsite cloud storage and data-processing centers?
Of course, in the United States, the president cannot unilaterally regulate AI with the stroke of a pen, let alone the issuance of a press release. Yet, curated policy briefs effectively crystalize the concerns that policymakers have about AI, those they would address right away, with a magic wand, if they could. If you were to read one, you might be surprised to learn that bread-and-butter health and safety issues still loom large, although those precipitated by workplace AI hew toward the mental or emotional and away from the immediate, physical, and catastrophic we alluded to earlier. The Benthamesque, managerial omnipresence facilitated by AI’s constant monitoring and surveillance, not to mention the pressure to meet AI-driven (or perhaps even AI-established) targets, can certainly raise stress, blood pressure, and burnout (Doellgast, O’Brady, Kim, and Walters 2023).
Aside from these physiological harms, unfettered and unregulated AI perpetuates bias and discrimination in ways that old-school heavy machinery did not (O’Neil 2016). Many learned about this initially in 2018, when Amazon publicly scrapped an otherwise secret AI recruiting tool that discriminated against women in the hiring process. Since large-language models (LLMs), one manifestation of AI, train on existing data, workers can easily be on the receiving end of seemingly objective hiring and promotion decisions predicated on a deep backlog of flawed human decision-making. The phrase “garbage in, garbage out” comes to mind: If past decisions were biased, and these decisions train the model, then we cannot be surprised when the model favors men. Likewise, without counterfactual data, how can a model know that the HBCU graduates who were not hired would have performed just as well or better than those Ivy League grads who were? Interestingly, a 2023 Zogby poll of more than 22,000 recruiters and others engaged in talent acquisition found that a third of respondents scored candidate engagement by running correspondence through LLMs (Employ 2023). Suffice it to say, workers might be unaware of exactly what sorts of data they themselves have passively and involuntarily fed these algorithms when they apply for employment or simply go about their work.
Before we continue, we should clarify the distinction between AI, LLMs, and algorithms writ large. While algorithms are specific sets of instructions for solving problems, AI refers to the broader field of developing systems that can perceive, learn, and reason in ways that mimic human intelligence. LLMs are a form of AI that harness algorithms as tools, but the algorithms themselves do not constitute AI; they are the underlying mathematical, logical, and computational processes that power these intelligent systems. For example, consider simple, Boolean, “if-then” style instructions predicated on logical values. LLMs are complex AI systems that leverage algorithms as building blocks to process and generate human-like text. These algorithms encompass various techniques, such as neural networks and language modeling, which enable LLMs to understand and produce coherent language—even if the system itself has no sentience or comprehension.
Between the black-box nature of an algorithm’s inner workings and the opaque ways in which workers themselves fuel them, the lack of transparency as well as data privacy concerns rise to the top of the AI policy agenda. In the United States, the Biden administration has called for keeping humans in the loop with clear governance systems, built-in and perpetual human oversight, and ongoing evaluation processes for workplace AI systems. Aside from transparency, the administration calls for worker-centeredness, providing employees with “genuine input in the design, development, testing, training, use, and oversight of AI systems for use in the workplace.” And even this nod to worker self-determination with respect to workplace technological change builds atop a more legalistic system of labor and employment rights as “AI systems should not violate or undermine workers’ right to organize, health and safety rights, wage and hour rights, and anti-discrimination and anti-retaliation protections” (White House 2024).
At least those navigating the aforementioned issues remain employed, as their bosses might be keen to remind them. Like so many technologies that have come before it, AI can automate tasks previously done by humans. Unlike previous technologies, the set of tasks it can tackle transcends the routine and repetitive (cf. Autor, Levy, and Murnane 2003) to include a broad swath of high-skilled, non-routine, white-collar work. Workers across sectors and occupations—not just those in manufacturing, customer service, and data entry—could see their terms and conditions degrade, particularly when policymakers leave it to market forces alone to prepare workers for AI-driven workplace transitions. Thus, the administration calls on employers to apply AI in augmentative rather than purely automative ways, using it to “assist, complement, and enable workers, and [to] improve job quality” and imploring employers to up-skill workers to facilitate employment transitions (White House 2024). Of course, the administration recognizes that it cannot expect let alone compel employers to hasten labor markets on their own. True policy fixes, the sort that address market failures, require the construction and maintenance of well-resourced institutions, a power that most representative democracies including ours do not bestow on any individual.
In the following pages, we dig deeper into the details of workplace AI regulation in the United States—part of this forum’s larger goal of analyzing this phenomenon on a global scope. That is, how are governments around the world approaching the regulation of workplace AI? Having just entered the era in which AI-powered algorithms influence managerial decisions, career trajectories, and economic outcomes writ large, policymakers around the world understand the time to establish regulatory standards is right now. After all, LLMs such as OpenAI's ChatGPT will not unilaterally drive outcomes for the firms that deploy them and the workers sweating alongside them. On the contrary, carefully crafted institutions or the void defined by their strategic omission will intervene in the otherwise straight-line relationship connecting workplace technological change to economic outcomes.
Global Perspectives on Regulatory Advances and Worker Influence over Workplace AI
The six essays that follow collectively paint a comprehensive picture of workplace AI regulation around the world. We begin with Mingwei Liu, Hao Zhang, and Yi Sui’s (2024) treatment of China, in part because of the PRC’s anomalous, unapologetically authoritarian stance on most economic and social matters. Perhaps not surprisingly, social stability remains a central focus of the government with respect to workplace AI regulation, particularly to the extent that the injection of AI into platform-mediated employment could foment unrest among these gig workers. Platform workers in China often operate beyond the bounds of formal employment altogether, depriving them of legal and established means for channeling redress. Consequently, the government will not risk a hands-off approach to workplace AI regulation, when doing so could allow for a direct connection between technological advance and social instability. The challenge, however, arises from China’s ambition to innovate: Since continued, fast-paced economic development redounds so directly to global economic competitiveness, Xi Jinping will not allow his government’s zeal for stability to materially slow its embrace of workplace AI. These dual objectives yield a policy strategy that Liu et al. label “strategic ambiguity,” meaning regulations that appear stringent and protective on the printed page but that grant authorities great flexibility and discretion to wield selectively.
This balancing act is a more concrete and detailed version of the policy dilemma we opened with above—the promotion of growth and the mitigation of harm, be it economic, social, physical. . . . Thus, it is not unique to China. Christine Bischoff, Ken Kamoche, and Geoffrey Wood (2024) explain that Kenya, indicative of those nations of the Global South seeking to find their place in the worldwide AI revolution, has also hosted a lively public debate on the ways to facilitate domestic innovation while at the same time at least nodding to concerns over worker exploitation. In the countries of eastern and southern Africa, where employers face little in the way of countervailing workplace power, they seem inclined to exploit AI’s potential to monitor and control some workers and to completely eliminate the work of others via new forms of automation. At the same time, the headline applications of AI that we read about in industrialized nations have engendered a wealth of new, low-end, “data janitorial” work. African workers have assumed much of this work creating the reliable data that developers rely on for training LLMs. Bischoff et al. note that, unfortunately, so many of these tasks—done remotely, informally, and yet easily monitored electronically—are unamenable to existing labor regulations. Ergo, a now familiar tension is emerging in these nations as well: Policymakers’ desire to secure some love from the global embrace of AI may lead governments to overlook the interests of their own citizens.
Similar pressures inhere in firmly established, industrialized democracies, too, as Sunghoon Kim and Seri No (2024) argue in their analysis of workplace AI regulation and worker resistance in South Korea. While the government has taken clear steps to regulate the collection and use of information, some initial choices appeared to privilege employers and, therefore, the development and use of AI tools, over those working with or alongside them. Only very recently have workers been able to mobilize around their rights as “data subjects” per se. Recent amendments to data protections first issued before the recent explosion of AI-powered algorithms now give workers the right to resist and even reject some managerial decisions on the grounds of data or processual opacity. Still, Korea’s techno-regulatory framework strongly resembles the European Union’s General Data Protection Regulation (GDPR), only with a more employer-friendly twist indicative of Korea’s more recent legacy of state-led economic development. For example, say Kim and No, the Korean law defines personal data more narrowly than the GDPR does and asks less of employers in terms of recordkeeping.
In her essay on workplace AI regulation in Germany, Didem Özkiziltan (2024) explains how the GDPR has come to play a central role in employment relations. Likewise, it must fit into existing institutions designed to protect workers while promoting safe workplace innovation. For example, the 2021 Works Council Modernization Act, aside from placing strict boundaries on employee monitoring and surveillance, meshes with existing requirements of the Works Constitution Act (WCA) and explicitly adds AI to its existing provisions. The WCA already required that works councils evaluate and approve of employers’ introduction and use of workplace surveillance tools. In fact, employers must provide these workplace-based representative bodies with comprehensive information, and regulators envisage works councils’ involvement in a broad span of issues related to the deployment of AI tools. Thus, on the one hand, the German case epitomizes a country with strong norms and well-established representative institutions, the sort which could allow individual firms and workplaces to self-regulate rather than rely on one-size-fits-all legislative solutions. On the other hand, Germany also offers a chance to consider the short- to medium-term costs of technologically induced economic adjustment. Özkiziltan does so through the lens of “Engels’ pause,” the label ascribed by Allen (2009) to Friedrich Engels’s mid-19th-century observation that the initial effects of the first Industrial Revolution were low wages, harsh working conditions, and broadening wealth gaps. In our view, this lens seems more empirically credible in the context of AI than the more commonly relied upon lens focused instead on technological unemployment per se. After all, nearly every job will be affected by AI, but long-predicted massive displacement has yet to materialize. And, shortening the duration of a modern-day Engels’ pause—or, to state it positively, hitting the play button on the narrative of widespread economic advancement—seems like a worthy policy goal.
The GDPR also shapes the regulation of workplace AI in the Nordic region. According to Anna Ilsøe, Trine Pernille Larsen, Christopher Mathieu, and Bertil Rolandsson (2024), Denmark and Sweden each combine their extensive welfare state apparatuses with their signature voluntarist systems of collective bargaining. This multilevel governance approach secures a stable, flexible, and dynamic regulation of the labor markets, with high collective bargaining coverage and union densities. The social partners in these countries respond to technological changes in tripartite, bipartite, and unilateral arenas, emphasizing collective agreements and collaboration between government, unions, and employers’ associations. Additionally, these countries have been skeptical of legal mandates and prefer multilateral, negotiated solutions, indicative of Nordic norms for labor market models and workplace regulation.
While the Nordic approach seems worthy of emulation, it relies on the firm entrenchment, deep involvement, and perceived legitimacy of social partners. In the final essay of the series, Antonio Rodrigues de Freitas Júnior, Letícia Ferrão Zapolla, and Paulo Fernando Nogueira Cunha (2024) suggest that the absence of such institutions challenges Brazil’s ability to regulate workplace AI constructively. The few laws enacted to date at the national level provide little in the way of worker protections. To fill those gaps, significant trade union involvement is needed. Yet, traditional unions find themselves too ill-equipped and under-resourced to bargain on behalf of nontraditional platform workers, laboring under algorithmic management. “Taskers” and food delivery workers show more affinity to the Brazilian equivalent of alt-labor organizations, nascent institutions at least as interested in broad, political, economic, and cultural matters as they are in so-called bread-and-butter issues. This echoes, albeit with some distortion, the US case. Conventional, American unions would love to include gig workers in their folds, particularly if policy changes would compel employers to bargain collectively over gig worker interests.
The Regulatory State of Play in the United States
White House policy briefs are one thing; black letter regulations are another. Nonetheless, emerging US regulation of workplace AI parallels the presidential administration’s rhetoric. By emphasizing workers’ interests as a major component of its AI industrial policies, it diverges from the traditional American hands-off and employer-focused approach to workplace and labor market policy. But the lack of robust workplace policy institutions, a federalized policymaking process, weak social partnerships between unions and management, and an activist judiciary shape the federal policy response to workplace AI.
In an unusual bipartisan push, the Biden administration led a federal legislative effort aimed at technological development and shepherded an AI-value-chain-focused industrial policy bill through Congress—the Creating Helpful Incentives to Produce Semiconductors (CHIPS) Act. The White House and multiple federal agencies in charge of data privacy, competition policy, anti-discrimination, and labor relations have issued substantive guidance memos about the utilization of AI in the private and public sectors. Alongside those federalized attempts, states and localities promote AI-related policies.
Surprisingly, and unlike previous tech scares, the White House has outlined a principled approach to developing federal AI policies, focusing ultimately on two main risks. First, are there AI-related harms to individuals and communities? Second, is there a risk of underutilization of AI in the private and public sectors? Not unlike Liu et al.’s (2024) analysis of the Chinese case, such a dual approach to AI is animated by both internal and external aspirations and pressures.
While other national cases in this forum struggle to adjust policies to compete for a bigger share in the AI value chain, China and the United States establish AI development as a core component in their struggle to lead the international economic order. Underutilization concerns, in both the private and the public sectors, are significant drivers in the Biden administration’s policy impetus (Office of Science and Technology Policy [OSTP] 2022). The federal administration often ties the push to mitigate harm to individuals, that is, protecting privacy or discrimination, to the risk of underutilization. Addressing harm increases AI “trustworthiness” and by that, increases demand for AI products (White House 2023). Addressing harms also clarifies legal ambiguities surrounding the application of existing policies and the reduction of legal risks. Additionally, addressing privacy and discrimination harms for workers and minority groups helps the Democratic administration cement internal support for its policies among political constituencies.
While the European Union and Korea compete with the United States for AI-related development and investments, they differ in that they have a centralized policy device to tackle information-related harms. While EU regulators and workers can build on the GDPR and their ongoing experience to extend it to novel technology policy problems, the United States lacks such a broad federal instrument and shared political experience. Similarly, unlike the EU policy experience, the capacities of the US federal administration and US workers’ organizations to engage with novel policy issues are under constant judicial scrutiny by a conservative judicial branch, generally hostile to federal workplace policy initiatives.
The United States also differs from its EU comparators by the weakness of its social partnerships and its eroded institutional environment for union–management relations. Unlike Germany and the Nordic examples, where multiple levels of workers’ input and voice are integrated into labor relations by law, the United States offers a singular, firm-based, adversarial framework. Unlike the Nordics, where strong collective bargaining institutions set baselines for AI implementation and manufacturing on multiple levels of labor relations, US private-sector union density stands at an ever-low 6% mark. Even more critically, labor law in no way guarantees even that small slice of the labor force a say in its employers’ technological adoption decisions. Once a union has been recognized as the sole representative of a workforce, the employer must bargain with the union “in good faith”—but only over those bargaining topics the law deems “mandatory” subjects of bargaining. Those include wages and other terms and conditions of employment. Technology-related decisions, like other matters of business strategy, are not mandatory, but merely “permissible” subjects of bargaining. Therefore, employers can and often do refuse to negotiate over these matters, (reluctantly) bringing only the “effects” of these decisions to the bargaining table, and only when the effects amount to changes in terms and conditions of employment. If those impediments were not enough, note that even these limited bargaining rights are not availed to most nontraditional workers, for example, contingent workers, contractors, and platform workers. Thus, examples of workers’ voice in developing and implementing workplace AI exist (Kresge 2020), and policy intervention offers workers some limited paths to have a voice. But US policy allows for only limited opportunities for workers’ voice and power over AI.
Finally, while China, the United States, and the EU compete for the top of the AI value chain and focus their policy responses around this pursuit, the Global South examples draw out policy responses aimed at capturing the opportunities and reducing the risks of being in a mid- to low-place in the AI value chain. Such a jurisdictional-based policy focus is highlighted by the lack of concern in US policymaking about the multitude of US-based “data janitors,” more commonly labeled “ghost workers” by American academics (Gray and Suri 2019). Such workers, toiling at the low ends of the AI value chain, are most susceptible to AI harm and exploitation but are generally invisible from the purview of US policy even when physically located in the United States.
Federal Workplace AI Policy
The most significant federal AI-related policy is the aforementioned CHIPS Act, which is designed to support the research and manufacturing of computer processors in the United States. Such manufacturing is considered the pinnacle of Biden’s emphasis on industrial policy targeted at capturing the lead in the international competition around AI development and applications. Substantively, the CHIPS Act provides financial incentives to the private sector, public–private partnerships, and research focused on developing and attracting manufacturing capacities. The CHIPS Act offers a range of targeted subsidies and tax cuts, promoting R&D in private companies and STEM education. As part of this investment in manufacturing capacities, the administration also allocates grants for regional economic development. Local governments and states now compete for such federal grants by assembling public–private partnerships focused on manufacturing, education, and infrastructure.
In addition to allocating funds for developing AI manufacturing capacities, the White House released three executive orders and several other policy documents focused on AI. All three executive orders instruct federal agencies to examine the application of their respective statutes to AI development. The executive orders also guide agencies in integrating AI into their own operations and investigating the effects of AI on unemployment and other labor market outcomes.
The latest executive order (EO), Executive Order on the Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence (White House 2023), deals with various issues concerning the regulation and use of AI in areas such as homeland security, biotechnology, and labor. The EO includes a long list of goals that ought to steer agencies in their development of AI guidance, including the “promoti[on of] responsible innovation, competition, and collaboration” and “a commitment to supporting American workers.” Section 6 of the EO directs the Secretary of Labor to collaborate with labor unions to create a “set of best practices” for employer use of AI that address concerns such as job displacement, protected activity, and transparency about data collection. Following the EOs, several agencies have issued guidance on the interpretation of their respective laws.
The Department of Labor (DOL), enforcing a range of workplace legislation from safety and health to wages and hours, has released only a few official documents regarding AI in the workplace. Internally, the DOL has made massive efforts to modernize its IT infrastructure (Ahluwalia 2022), integrating AI into its daily operations. The DOL’s Use Case Inventory (U.S. Department of Labor Office of the Assistant Secretary for Administration & Management 2024) provides insight into how the DOL relies on AI to perform its duties, for example, using AI to read medical documents to determine causal language, flying drones to inspect unsafe working conditions, and deploying LLMs to generate information on workplace injuries for unemployment insurance reports.
The National Labor Relations Board (NLRB), entrusted with enforcing the National Labor Relations Act (NLRA), is primarily concerned with AI’s implications for workers’ right to organize. In 2022, the NLRB’s General Counsel, which prosecutes violations of the NLRA, issued guidance clarifying the scope of the NLRA’s protections against surveillance and automated management tools. While surveillance systems do not have to incorporate AI to violate the NLRA, the memo argues that AI offers employers the capacity to collect and analyze significant amounts of data, which may enhance the impact of surveillance and chill organizing.
The Equal Employment Opportunity Commission (EEOC), enforcing various workplace anti-discrimination laws, has launched the Artificial Intelligence and Algorithmic Fairness Initiative (U.S. Equal Employment Opportunity Commission 2024) to clarify how anti-discrimination laws align with employers’ use of AI. The EEOC is concerned primarily with the impacts of AI assessment tools on people with disabilities who qualify for reasonable accommodations under the Americans with Disabilities Act (ADA). According to the EEOC, it is unclear how the legally mandated accommodation is calibrated into assessment tools and protocols; it therefore recommends that employers use assessment tools that can take employers’ obligation to accommodate disability into account (U.S. Equal Employment Opportunity Commission 2022). In addition, the EEOC has clarified that both the employer and a third-party AI software provider might be liable for discrimination against workers or applicants.
Of note is that agency guidance and opinions are susceptible to court scrutiny and a change of federal administration. In at least one recent case, a circuit court rejected the NLRA’s broad read of the impact of surveillance on organizing. As agencies move to enforce their interpretation of their respective laws, pushback through courts is expected. Such pushback will likely join the existing attack on federal agencies’ constitutionality and authority.
State- and Local-Level AI Regulation
In the past decades, US states and localities have been active in advancing workplace policymaking. From minimum wage hikes to anti-discrimination and scheduling protections, local government pushed workers’ protections to new levels (Galvin 2021). Such policies result from workers’ political gains in numerous blue cities and states (meaning a Democratic majority). Still, local policies are susceptible to both hostile courts and preemption by states (in the case of local governments) and by the federal government (in the case of both cities and states) (Johnson 2021; Racabi 2021). As such, while pursuing the goal of AI-driven economic growth, cities and states offer some room for workplace policy entrepreneurship and experimentation. Early examples of such initiatives follow.
Many states, such as California and Illinois, have general data and biometric data protections that apply to workers and a host of anti-discrimination regulations that apply to most employment actions, including actions made by AI. However, while multiple states are pushing for AI regulations, Colorado is the only state with a law that creates substantive rights for individual workers for information, correction of data, and appeal (Marr and Williams 2024). In May 2024, Colorado enacted SB 24-205, a dedicated AI anti-discrimination law. The Colorado law affects AI-facilitated decision-making in education, employment, financial services, health care, housing, insurance, and legal services (“High-Risk” AI Systems). It prohibits discrimination based on age, color, disability, ethnicity, genetic information, English proficiency, national origin, race, religion, reproductive health, sex, or veteran status (SB 24-205 Secs. 1(A), 3).
The Colorado law then defines two main regulated actors: Developers and Deployers of regulated AI systems. Starting on February 1, 2026, Developers are obligated to perform a host of risk and benefit assessments of their AI systems and to collect and disclose various information regarding the development, data, and outputs utilized in the AI system to Deployers of the AI system and the Colorado Regulator (Sec. 6-1-1702). Deployers of regulated AI systems, alongside similar disclosure requirements, must deploy a risk-management policy and an impact assessment calibrated to the particular nature of the AI system and the risks such a system entails. In addition, Deployers must notify people affected by the application of a regulated AI system on their case, including the data being used, and allow individuals to correct or add data and appeal adverse decisions “if technically feasible, allow[ing] for human review” of the decision (Sec. 6-1-1703).
On the local level, New York City (NYC) is a pioneer with an existing law regulating workplace AI decision-making. Since 2023, NYC has required employers and employment agencies to independently audit AI decision-making about hiring and promotions and disclose the use of such tools to applicants and workers (Local Law 144, 2021). NYC also protects fast-food workers from termination following specific pervasive electronic surveillance methods in its fast-food sector regulations. The effectiveness of those local measures is questioned, however, and statewide efforts are underway to fill the gaps in those measures (Marr and Williams 2024).
Alongside those concrete, existing examples, states and localities have been pushing their versions of a designated AI bill of rights, mainly focusing on the same areas as the Colorado law (e.g., housing, health care, employment, and so forth) and awarding the subjects of AI decision-making an equivalent set of rights: information, disclosure, and an affirmative right against discrimination (Bernhardt and Pathak 2024).
A consistent theme in US AI regulations is their focus on individual harms, rights, and institutions (OSTP 2022). Considering the broad preemption regime of the NLRA, which prohibits states from regulating private-sector union–management relations, states and localities are likely locked into this individual harms, rights, and institutions framework. However, some complementarity with unions’ negotiations is available to policymakers. For example, Hollywood writers and actors, who recently achieved collective bargaining clauses protecting members from certain AI risks (Litwin 2023), are lobbying for state laws that would cement those wins in the policy front. The New York Digital Replica Contracts Act, which prohibits unauthorized replication of an actors’ voice or appearance, and New York Senate Bill S7422A, which excluded film productions that utilize AI to replace employees from a state tax-subsidy scheme, augment the film industry unions’ collective bargaining wins. Further study of unions’ advocacy and utilization of AI policies, and their relations with other social partners on the federal, state, and local policy fronts, is a promising front for US labor relations research.
Conclusion
The US case demonstrates an early perspective on the emergence of AI workplace regulations across both the federal and the local levels. While the federal government is pushing for an industrial policy that would push the United States to the top of the AI production and manufacturing value chain, it seeks to do so in a way that would be trustworthy and equitable for consumers and workers. The federal push is limited by a lack of strong workers’ voice institutions, however, which is a product of low density and weak labor law that does not compel employers to bargain over technological choices. Like the federal government, states and cities are focused on individual harms and face fewer obstacles on the path to institutional reform, but they oftentimes lack capacity and are limited in their ability to regulate and promote collective rights. Yet, some unions, as demonstrated in the short New York case, can offer policy initiatives that would complement traditional collective voice and power mechanisms.
In a sense, the biggest concern in the United States is one that arises thematically across all the essays in this forum. These are early days for workplace AI, and thus, even earlier days for AI regulation. Regulators must strike a delicate balance, protecting workers without letting them get left behind. And, just how they do this depends on where their worker-constituents operate in the value chain. But, perhaps even more critically, policymakers across all of these countries and regions want to stake a regulatory claim; only then can they comfortably adjust their approach as the technology’s capabilities evolve and advance. In the essays that follow, we invite you to see just how these efforts are faring across the globe.
