Report | AI

Federal AI legislation: An evaluation of existing proposals and a road map forward

Download PDF

The introduction of new artificial-intelligence-based technologies has generated front-page headlines and grabbed the attention of consumers and policymakers recently. While related technologies have been in use in workplaces for many years, new commercially successful products have spurred greater discussion about the impact of artificial intelligence (AI) on our economy and society. Similarly, new research and reporting have highlighted the direct experience of workers who use or are subject to these technologies in industries ranging from warehousing and manufacturing to health care and retail services. In response to growing attention and concerns with the labor market impacts of AI technologies, policymakers at nearly every level of government have published principles, issued new guidance, and introduced legislation on a range of AI-related issues—including data privacy, employer disclosure practices, and auditing requirements.

This report provides a brief overview of what is known so far about the economics of AI, highlights some accounts of how it is being deployed in anti-worker ways across different industries, offers a landscape analysis of federal legislation, and then presents policy recommendations on what federal legislation should be aiming to achieve given the latest research on the likely impacts of AI on workers and our economy. A key theme of the policy recommendations is to keep sight of the broader economic and institutional contexts in which AI might be deployed and to avoid tunnel vision in crafting policies that are too narrowly focused on the latest tools used by employers (rather than the underlying harmful practice).

The deployment of AI-powered technologies in the workplace has been associated with an array of worker experiences and outcomes that warrant concern, including discrimination, unsafe working conditions, seemingly arbitrary disciplinary action or discharge, among others. This has led many policymakers and advocates to put AI on the top of their priority list for drafting new legislation or regulations meant to protect workers. Policymakers have generally focused on assessing how the introduction of AI might degrade the on-the-ground experience of workers. While these assessments are illuminating, they have often steered policy responses toward solutions that are tailored to current forms of AI technologies or use cases. This approach, however, leaves policymakers in the position of potentially failing to keep up with other tools and practices available to employers that result in the very same outcomes (especially as technologies continue to advance) and runs the risk of falling short of providing meaningful long-term benefits to workers.

For years, employers have used algorithmic or automated systems in ways that harm workers—including through discrimination, diminishing workplace safety and privacy, and limiting decision-making power. And for decades, employers have been empowered to use garden-variety management practices that do all these things. Policymakers should rightly be alarmed by the experience of workers today, but they should look beyond the latest tool businesses are using to achieve these outcomes and instead look deeper to the underlying conditions that enable these results.

At the most simple level, the impact of AI technologies on workers is a function of the balance of power between businesses and their workers. If there is balance in the labor market because policy has empowered workers, then most new technologies (including AI) will be steered into generating productivity gains that can be shared across the economy. If the labor market remains unbalanced and workers disempowered, then lots of new technological tools will instead be deployed in zero-sum ways to keep wages suppressed and the incomes of capital owners and managers high.

The rest of this report builds to a set of recommendations for how policymakers should approach empowering workers in the age of AI. First, we review the likely economic impacts of AI technologies. Second, we provide a landscape analysis of existing federal legislative proposals. Finally, we outline three key pillars of a worker-centered policy strategy in the age of AI that aim to achieve a simple objective: increase the ability of workers to meaningfully engage their employers in how AI technologies are deployed in the workplace. These pillars are: 1) expand and expedite pathways to collective bargaining, 2) reduce AI-specific barriers to worker voice and strengthen employment protections, and 3) increase workers’ ability to leave employers who create exploitative conditions with AI (or other technologies and practices) and ease job transitions.

Labor market impacts of advancements in AI

There has been a great deal of speculation and attention given to the potential labor market impacts of new AI technologies over the last few years. In particular, many people have voiced worries about mass job loss and the impact of technologies on skills required for various occupations.1 Others have raised concerns about AI undermining core workplace rights and altering conditions of employment.2 In this section we identify what we believe are the known labor market impacts of AI. These evidence-based outcomes should serve as guideposts for evaluating current legislative proposals and inform additional policy recommendations.

Aggregate labor market impacts: Employment effects and employer demand for skills

While anecdotes and limited data analysis have fueled conversations about mass job loss, it’s important to consider a longer-run assessment of the impacts of technology on the labor market, drawing on a rich body of economic research and measurable outcomes during prior periods of technological advancement. 

Earlier this year, Bivens and Zipperer (2024) published a thorough assessment of AI’s likely impact on the labor market and workers, given the evidence from past waves of major technological progress. Their analysis highlights that technological advancements and subsequent productivity gains are highly unlikely to lead to mass joblessness.

Concerns that technological advancements will create significant job loss rest on the basic theory that AI technologies will increase productivity such that firms will use less labor, resulting in mass layoffs and a rise in unemployment. However, this theory muddles firm-level experiences with aggregate outcomes and glosses over a crucial piece of the equation: whether there will be other jobs in the labor market available—including those created by technological change—for workers who may be laid off from an individual firm.

Unemployment increases when the potential output of the economy (how much could be produced if nearly all of the labor force were employed) exceeds aggregate demand (total spending by households, businesses, and government). Thus, even if AI-based technologies result in an increase in productivity, whether this will result in an increase in unemployment depends on if aggregate demand doesn’t similarly increase. Bivens and Zipperer (2024) further note that fiscal and monetary policies enable aggregate demand to be boosted much more quickly than potential output can move, allowing policymakers to minimize the magnitude and duration of any disconnect between potential output and aggregate demand.

Relatedly, a review of the relationship between productivity growth and unemployment highlights that productivity gains do not occur at the expense of rising unemployment. In fact, Bivens and Zipperer (2024) find that “fast productivity growth is associated with lower average rates of unemployment across business cycles.” Moreover, their analysis shows that faster productivity growth does not hamper the pace of falling unemployment over a business cycle. In short, there is no compelling evidence to support the theory that AI-induced productivity gains will result in mass workforce displacement. 

Beyond aggregate employment effects, there has been a corollary concern among some policymakers about how businesses’ use of AI technologies may impact the demand for workers without a college degree. That is, as more complex technologies are developed and deployed in the workplace, will firms hire more highly skilled individuals for jobs that previously required lower skills? Here too, the theory for why this may occur is straightforward, but a dig into the data reveals why this is unlikely to occur in practice.

First the economic theory: An increase in use or advancements in technology will lead to an increase in demand for more highly skilled workers relative to workers who don’t have the skill set to use these technologies. Rising demand for higher-skilled workers will, as the theory goes, increase wages among those relative to lower-skilled workers, increasing income inequality. In general, economists have often used a college degree as a proxy for workers who obtain the skills necessary for new technologies. 

As Bivens and Zipperer (2024) discuss in greater length, a review of the data shows that this theory hasn’t played out in practice over the last 20 plus years. Since 2000, there has been little to no change in the college wage premium despite significant technological advancements. In fact, over the last few years the college-to-high-school wage premium has fallen, resulting in a decline in wage inequality between higher- and lower-skilled workers. Therefore, there is little evidence to suggest that continued advancements in AI will trigger a greater need for “upskilling” to counterbalance decreasing demand for lower-skilled workers (Bivens and Zipperer 2024). 

In the period since 1979, there has been a consistent shift in labor market power away from typical workers and toward corporate managers and business owners, resulting in a redistribution of income. However, as Bivens and Zipperer’s analysis shows, a close review of economic research does not support the notion that technological advancements over this period have been the causal driver of worsened outcomes for workers. While technological change may have coincided with rising income inequality, it was itself not a headwind for workers or our economy. Conversely, under the right conditions—more equal distribution of market power between businesses and workers—advances in technology can and have resulted in improved outcomes for workers.

It’s not difficult to see how improvements in technology will help spur greater economic growth by making production processes more efficient. The question then becomes whether workers will be able to share in the gains that flow from new advancements in technology. The answer hinges on the relative labor market power of workers. Given today’s significant imbalances and the erosion of core labor rights, it is unlikely that workers will share fully in the economic gains brought about by AI technologies or any other driver of productivity growth. 

Industry-, firm-, and worker-level impacts of AI 

Beyond aggregate-level impacts, many studies and reporting chronicled the experience of workers using AI and how the use of these technologies have played out within specific industries or occupations. Below we discuss the types of AI technologies that are being used across three industries—warehousing, call centers, and health care; how the specific manner in which they have been implemented has exacerbated negative outcomes for workers; and the efforts workers and their representatives have taken to mitigate these effects.

AI in warehousing

The use of AI technologies in the warehousing industry is perhaps among the most publicly recognized. Over the last few years, there has been extensive documentation of the use of technologies in this industry to set productivity targets and task management, particularly among large retailers like Amazon and Walmart.3 For example, handheld devices track individuals’ pace of work including the number of packages they may be scanning in an hour, error rates, and time between scans (Bernhardt, Suleiman, and Kresge 2021). These data are then used by companies to establish productivity quotas or inform disciplinary decisions. Additionally, robot carts can be used to direct where workers should go in a distribution center, identify which products to move, and set the pace of work (Bernhardt, Suleiman, and Kresge 2021). 

While this is an industry with historically higher injury and illness rates than national averages, these incidents have ticked up in recent years.4 To be sure, not all injuries are a result of using AI technologies, but it is clear that the manner in which these technologies are currently being used in many establishments is associated with severe injuries.

The National Employment Law Project (NELP) found that 1 in 15 Amazon employees experiences a recordable injury each year in an analysis of Occupational Safety and Health Administration (OSHA) injury reporting records for establishments with more than 1,000 workers (Tung, Marquez, and Sonn 2024). Amazon’s injury rate has consistently been higher than the national average and even exceeds other industries with notable workplace hazards such as coal mining, forestry, and logging (Athena Coalition et al. 2020). As the researchers note in an earlier study of OSHA’s records, the most common injuries are to “workers’ backs, shoulders, knees, wrists, ankles and elbows. These types of injuries are often caused by workers assigned tasks involving ergonomic hazards including forceful exertions, repetitive motions, twisting, bending, and awkward postures” (Athena Coalition et al. 2020).

Almost all cases (95%) result in employees missing work or being assigned to different job duties; these injuries can have lifelong impacts and associated problems (Tung, Marquez, and Sonn 2024). The report goes on to explain that “the high rates of serious injury at Amazon are directly attributable to the way that the company manages its workforce using intensive surveillance, automated discipline, and constantly changing quotas generated by algorithms” (Tung, Marquez, and Sonn 2024).

Injuries in the warehousing sector are not unique to Amazon, though research by NELP indicates rates of injuries are far higher at their facilities. Recognizing the prevalence and severity of injuries in the warehousing sector caused by a range of practices, the Occupational Safety and Health Administration launched a national emphasis program (NEP) in 2023 to address comprehensive hazards in this sector (OSHA 2023a). Notably, the NEP’s inspection procedures included an ergonomics screening. Specifically, during interviews with workers, while reviewing injury logs, and site walkthroughs, OSHA inspectors will assess whether workers are exposed to ergonomic hazards and if so, then expanding the scope of investigation (OSHA 2023b). Importantly, in the absence of a specific ergonomics standard, OSHA can only issue citations related to ergonomic hazards against employers under the general duties clause of the OSH Act (OSHA 2024).

In response to poor working conditions, workers and advocates have sought solutions through labor organizing efforts and state-level legislation. While strides are being made in certain states like California, Minnesota, New York, and Washington, new organizing efforts at individual establishments have been arduous. For example, reporting has shown that Amazon may be using electronic monitoring and productivity quotas as a means to retaliate against employees engaged in union organizing. Specifically, in one case at a Kentucky fulfillment center, an employee filed a complaint with the NLRB claiming that Amazon used his failure to meet specific performance targets as pretext to retaliate against him for leading a union organizing campaign (Rosenberg 2022). While Amazon employees at a Staten Island fulfillment center successfully voted to join a union in 2022, they have not yet secured a first contract (Chapman and Hadero 2024).

In an analysis and discussion of injury rates at Amazon, Julia Lang Gordon (2021) explains how workers’ diminished market power—particularly, limited alternative employment opportunities and the prospect of being fired without cause—has enabled “Amazon to push its employees to the physical brink while facing little to no repercussions.” At establishments where workers are able to secure representation by a union and a collective bargaining agreement, there is a clear pathway to improving working conditions, including those that are undermined by AI technologies. It would be exceedingly unlikely, for example, for a unionized employee to be fired for failing to meet unrealistic productivity quotas that if met, would require engaging in unsafe practices; workers would similarly have a clear process to protect basic rights like accessing the bathroom and taking rest breaks.

The warehousing industry is a clear example of how workers’ lack of bargaining power has directly limited their ability to engage businesses in how technologies are used in the workplace and prevent unsafe practices that have resulted from their use. 

AI in call centers 

The impact of electronic monitoring and management on working conditions is not unique to the warehousing sector. In call centers, AI technologies have been deployed to replace, assist, and manage workers’ activities. For example, automated systems are being used to screen calls and route them to employees; chatbots are being used to provide workers with prompts and answers to callers’ questions; other systems are generating notes and transcripts of employees’ calls with customers, identifying patterns in behavior or deviations from call scripts; and other technologies are providing real-time management to employees, including prompts related to their tone, pace of speech, and cues to signal specific emotions (Bernhardt, Suleiman, and Kresge 2021; Doellgast et al. 2023).

In a recent survey of call center workers, researchers explored how the use of these technologies impacted workers’ outcomes. While the experience with these technologies varied, most workers held negative views of AI management tools. About two-third of workers reported that automated monitoring systems made their jobs more stressful and did not feel that monitoring or coaching systems increased fairness on the job (Doellgast et al. 2023).

Among the most concerning impacts is the potential for real-time recording systems to generate performance evaluations, and then inform disciplinary decisions and firing without employees having an opportunity to understand what is driving these decisions, the accuracy of the data underlying them, or avenues for appeal. In a written statement for a U.S. Senate forum on AI, Ameenah Salaam (2023) of the Communications Workers of America describes how CWA workers are experiencing this very dynamic: 

These systems are often ineffective and have negative impacts on the workplace and the quality of service our members can provide. We’ve heard from workers of color about discriminatory bias from systems that purport to judge expressions of emotion, like empathy; workers also say the systems do not recognize certain pronunciations and styles of speech. Agents report that the scripts enforced by these systems slow down the work of helping customers and often advise wrong solutions that may violate company policies, creating a situation where agents can face discipline for following the system’s prompts.  

As with any workplace, the presence of a union enables workers to engage their employers and address how specific business practices affect job quality and their well-being; this is particularly true for call centers where employees have long grappled with electronic surveillance and monitoring. Research on the impact of unions in call centers has found that unions not only increase worker well-being but are able to successfully negotiate over how information gathered from performance monitoring is used (O’Brady and Doellgast 2021). Sean O’Brady and Virginia Doellgast (2021) write, “unions place a strong emphasis on challenging ‘discipline-based’ performance management practices and encouraging more developmental ones focused on training and development, as well as in establishing more fair and transparent processes for evaluating and rewarding performance.” In fact, in their study on the impact of unions on key worker outcomes, the authors found that union activities were positively associated with developmental performance monitoring and a greater perception among workers of the fairness of performance metrics. Overall, the authors found that union engagement improved worker well-being (measured by emotional exhaustion) (O’Brady and Doellgast 2021). Similarly, as Aurelia Glass (2024) has written, the CWA has over the last three decades secured contracts with major companies, including AT&T and Verizon, that have placed limits on the frequency of monitoring employees calls and the ways in which recordings can be used to inform disciplinary action.

Thus, it is clear that to the extent workers and policymakers are concerned about how AI management systems or performance metrics will be used and the fairness of the underlying AI technologies, unions have a clear track record of being able to address these very issues.

AI in health care

In the health care sector, workers have raised significant concerns over how AI-related technologies have impacted staffing ratios, limited practitioners’ discretion in patient care, and resulted in incomplete or incorrect care plans. In a recent survey of registered nurses, more than 50% said their employers utilized an algorithmic system to analyze patients’ health records to assess patient care needs (NNU 2024). The survey, conducted by the National Nurses Union (NNU), also found that a growing number of shift handoff reports from one nurse to the next was partially or wholly generated by AI technologies. And between 30% and 40% of nurses whose employers use AI systems are not able to override or correct care assessments or outcome predictions associated with discharge recommendations (NNU 2024).

The NNU’s research alarmingly finds that there are notable deficiencies in the accuracy of AI systems in use. For example, nearly 70% of nurses who use algorithmic systems to assess patient acuity reported that their own assessment did not match that of the computer. Similarly, nearly 50% of nurses who receive AI-generated handoff reports said their own assessments differ from the information provided to them (NNU 2024). 

Like warehousing, call centers, and other industries where AI technology is being deployed, the use of AI in the health care setting is exacerbating long-standing issues that health care professionals have bargained over, including staffing ratios, and discretion in patient care. While NNU and others are ringing alarm bells for policymakers and the public on the impact of AI on nurses and patient care, the impact of AI on health care professions goes well beyond nurses. Physicians and others who are less commonly represented by labor unions may face greater challenges in raising awareness of these issues and mitigating harmful outcomes.

The three industries discussed above, in addition to research on the use of AI in other industries and occupations, highlight three general themes that should inform policy evaluation and development. First, AI technologies are often the latest iteration of prior technologies and management strategies. While the use of certain AI technologies is becoming standard practice in certain industries and occupations, there remains significant variation in use across firms. Second, there is a lack of transparency around when technologies are being used, how they function, and how they inform employment-related decisions. Third, and perhaps most importantly, under certain circumstances AI-based technologies are undermining existing labor and employment rights. It is important to note, that these three broad findings and trends do not focus on the inherent impact of AI technologies but rather the impact of businesses using these technologies in the context of today’s labor market—most crucially the current balance of power between workers and businesses. Across a range of sectors, unions are negotiating over employers’ use of AI in the workplace and have secured crucial wins for workers (Glass 2024). Given weakened labor laws, though, it is exceedingly difficult for most workers today to successfully secure representation and bargain over AI and other issues.

Given the aggregate-level labor market impact of AI as well as the variety of ways AI technologies are being used across industries and firms, the objective of federal legislation should be to increase the ability of workers to meaningfully engage their employers in how AI technologies are deployed in the workplace.

Landscape of federal legislation 

There is a bevy of federal legislation, principles, and road maps intended to curb negative impacts of AI on workers. In the appendix below, we provide a complete review of these bills, highlighting the most common policy interventions. After reviewing these proposals, it’s clear that few bills include comprehensive solutions for workers in a period of AI advancement. For example, many bills include new employer disclosure requirements, ensuring workers are aware of when certain practices like workplace surveillance and data collection are occurring. However, given workers’ lack of bargaining power and low unionization rates, it may be challenging for the typical worker to adequately leverage this new information on their own to change their employers’ practices.

Notably, one proposal—the Warehouse Worker Protection Act—takes a novel approach by aiming to address head-on unique barriers to organizing in the warehousing industry. The bill does this in two ways. First, it limits the ability of employers to fire workers for failing to meet productivity quotas that haven’t been previously disclosed to the employee and requires greater transparency of data disclosure to allow employees and their representatives to assess whether employment actions based on performance monitoring are consistent across employees. Second, the bill would amend the National Labor Relations Act and the Occupational Safety and Health Act to strengthen the ability of these agencies to enforce workers’ rights. But even here, the amendments are narrowly tailored.

Most of the proposed bills to date aim to address concerning issues workers are facing—discrimination, invasion of privacy, lack of recourse for hiring, disciplinary or other employment actions, and safety risks. But unless AI-specific proposals are paired with solutions to address workers’ bargaining power, these AI policies will fall short of their very objectives. Congress may soon be caught in a game of whack-a-mole, chasing the latest use case of AI systems as employers adjust practices in response to new legislation and technologies, all while working conditions continue to deteriorate.

Policy discussion and recommendations

In the prior sections, we discussed the likely impact of AI on the labor market as well as the current experience of workers who use or are subject to AI-powered technologies. Taken together, these economic realities have led policymakers to question what policy interventions are needed during a period of AI advancement. However, we believe the success of any AI policy will hinge on its ability to address the underlying conditions that have, for decades, been enabling businesses to use the latest tool available to them—in this case, AI technologies—to undermine worker power and erode workplace conditions and outcomes. Fundamentally, the impact of AI on workers is not solely a function of the design of the tool itself but instead the institutions and policies that shape how it is implemented in firms across America.

Workers should have a voice in how policies are implemented in their workplaces. But decades of weakened bargaining power and increased reliance on employers for basic necessities such as health care and retirement have made it exceedingly difficult—if not impossible—for most workers to bargain over how AI technologies are being used in the workplace, let alone walk away from poor working conditions. These broader dynamics have led to the very outcomes policymakers are rightly concerned about when it comes to AI: diminishing workplace safety, discriminatory hiring, limited decision-making power, lack of data privacy, and disciplinary actions without explanation or appeal.

Therefore, any set of AI policies must first address the systematic erosion of core labor rights and meaningful exit options for workers in exploitative job situations. Conversely, to the extent policymakers aim to construct AI-specific policies, they need to do so strategically. This means assessing whether proposed solutions to common outcomes for workers who use or are subject to AI will 1) meaningfully remove unique impediments to the exercise of worker voice on the job or 2) create new protections that will be realized by workers beyond today’s latest version or use of AI systems. Policymakers may be tempted to craft solutions that are narrowly focused on addressing the consequences of the use of AI technologies today. But this approach runs the very real risk of crafting solutions that quickly become ineffective as employers’ practices change or technologies evolve. Instead, a wider lens approach to creating standards and protections that happen to also stand up against employers’ current use of AI technologies will be more likely to meaningfully improve worker outcomes for years to come.

Indeed, Congress has long had to balance this dynamic—deciding whether to regulate an outcome or the mode by which an outcome is realized—when designing some of the nation’s most fundamental employment protections. For example, when Congress passed the Fair Labor Standards Act, legislators debated whether to explicitly include a list of the common employment structures that they believed businesses would use to evade coverage under the FLSA such as piece work, off-premises work, and commissions, among others. As Kati Griffith (2019) writes in a historical analysis of the FLSA:

Congress eventually rejected this much-discussed list of specified devices of evasion in favor of a much broader and more flexible concept of employment. It provided definitions that could adapt with the times and adapt to new “devices.”…By moving away from a specific list of tools of evasion, Congress gestured that it did not want to slide into an endless game of whack-a-mole, to preempt different business structures and strategies that might emerge to sidestep the FLSA’s coverage. 

While not directly analogous to the choices Congress faces today, the legislative history of the FLSA is nonetheless informative as Congress considers how and to what extent they should regulate AI in the workplace. If AI-related outcomes highlight weakness in existing rights (like health and safety) then the foundational statutes, regulations, and enforcement efforts should be strengthened to ensure the outcomes arrived at by the use of AI are indeed protected as intended. Similarly, to the extent AI-related outcomes have shone a light on new or unlegislated risks to workers, then Congress should address the issue broadly; for example, issues related to data privacy should be addressed in a manner that ensures clear lines of privacy in the workplace are drawn for employees and personal information—no matter how it is obtained—is not used for illegitimate purposes.

Below we discuss three pillars of a worker-centered AI policy strategy that draws on the themes discussed above with the aim of achieving a straightforward objective: increasing the ability of workers to meaningfully engage their employers in how AI technologies are used in the workplace. 

Pillar one: Expand and expedite pathways to collective bargaining

Much of workers’ experiences with using or being subject to some form of AI technology today is mostly a function of low unionization and weak bargaining power, not of the current state of technology. Relatedly, efforts to increase transparency around the use of AI technologies (discussed below) will only be effective at catalyzing change if workers and their representatives are able to act upon that information. Therefore, a key pillar of any worker-centered AI policy strategy is expanding and expediting pathways to collective bargaining. The most powerful way to do this would be for Congress to pass the PRO Act and the Public Service Freedom to Negotiate Act (PSFNA). However, this much needed fix to our labor laws has languished in Congress for far too long. 

While considering policies to support workers and create good jobs in an age of AI, policymakers should also look closely at other ways to restore collective bargaining rights and increase bargaining power of workers, even if it is short of the more comprehensive solutions of the PRO Act and PSFNA, or limited to AI-related issues.5 In many other countries, sectoral bargaining is used to reach industrywide agreements on core conditions of work. In the United States, though, sectoral bargaining has not taken hold, but the Clean Slate for Worker Power project and others have put forward a rich set of policy recommendations on how to advance sectoral bargaining in the U.S. (Clean Slate 2021; CLJE Lab 2024a). For example, tripartite boards composed of representatives from employers, workers, and the public could be convened to negotiate new standards for specific industries or issues (Madland 2019). Congress should consider whether emerging dynamics related to AI may be ripe to address through sectoral bargaining, which is particularly useful in sectors in which there is already strong union representation or, conversely, high levels of monopsony power. 

Pillar two: Reduce AI-specific barriers to worker voice and strengthen employment protections

Workers in the United States—particularly nonunion workers—are far too constricted in how they can express voice over how their workplace is organized and run. One key impediment is the opaque management practices and production processes in workplaces. AI—a highly specialized technology that very few understand deeply—threatens to make expressing voice over workplace practices even harder.

Policymakers have, therefore, focused much of their attention on using legislative interventions to increase firm-level awareness and knowledge of technologies that are in use. However, it’s important to realize that legislative solutions that fill the knowledge gap that results from weak bargaining power is only part of the equation. In other countries with strong labor unions, employees have been able to use their market power to not only extract a clear understanding of what technologies are being used, but to use that information to further engage businesses to bargain over how technologies should be used. It is this latter part of the equation where current federal proposals seem to fall short. If, in the absence of adequate worker bargaining power that would allow workers to effectively demand transparency from employers on their own, Congress is attempting to step in and require disclosure of technological practices, then these proposals should be created in a way that maximizes the ability of labor unions, worker advocates, and nonunion employees to realistically use that information.

Therefore, policies that include new employer disclosure requirements to workers should also create a public repository of businesses’ disclosure reports. Greater public access to this information will reduce barriers for labor unions, worker advocates, and researchers to access the information, enabling these entities to identify and take action on industry trends and worker outcomes. Similarly, proposals that create new data privacy protection for workers, which often include a right for workers to receive and review their data, should also consider how workers could realistically understand and act upon this information. Other experts, for example, have suggested the use of AI monitors in the workplace, which would help create a venue for workers—union and nonunion alike—to better understand disclosures provided by employers and engage with them directly on complex issues (CLJE Lab 2024b).

Beyond increasing transparency related to the use of AI, there are a number of existing labor and employment laws that limit its use. While these existing laws give workers some voice and power in their workplaces, it is crucial to ensure that the laws are not subverted by the deployment of AI in workplaces.

The EEOC, NLRB, and DOL have all issued various forms of guidance clarifying the applicability of relevant laws to the use of AI technologies. The EEOC has multiple resources for employers related to the use of AI in hiring-related decisions and is actively engaged in enforcement actions against businesses that have violated the law (EEOC 2023). Similarly, the NLRB general counsel has issued a memorandum identifying how the use of surveillance and monitoring technologies may violate core employee rights under the NLRA (NLRB 2022). Finally, the Department of Labor’s Wage and Hour Division has issued a Field Assistance Bulletin to their investigators, clarifying how AI technologies may result in violations of minimum and overtime standards, as well as employees’ rights under the FMLA, and other laws (DOL 2024).

Looking ahead, policymakers should first and foremost ensure enforcement agencies are adequately funded. While the size of the workforce and complexity of businesses continue to grow, for too long agencies’ enforcement capacities have not kept up. Enforcement agencies will be limited in their ability to ensure employers’ use of AI technologies does not run afoul of existing laws if Congress does not provide them with adequate funding. Relatedly, Congress should consider what new requirements might aid in the compliance and enforcement actions under existing laws. For example, policymakers should leverage the authority and expertise of existing federal entities, such as EEOC, NIST, or the FTC, to standardize auditing frameworks and tools that can be easily used for pre-deployment and ongoing testing. The first-order objective of these audits should be to identify and eliminate outcomes that violate existing labor and employment laws. Additionally, as Congress considers new proposals, whether they are disclosure requirements or new employment standards, policymakers should take steps to ensure there are avenues for workers and their representatives to more fully participate in the enforcement process, including a private right of action.

Additionally, policymakers and federal agencies should consider how to more fully leverage existing authorities to address workplace outcomes that have been exacerbated by the use of AI technologies by issuing new guidance and regulatory standards. For example, OSHA should consider which industry-specific standards would better protect employees from workplace injuries that can arise when certain AI technologies are deployed in a specific manner; NIOSH should study the mental health impacts of specific working conditions and practices, including those with AI technologies, creating the evidentiary basis of potential OSHA standards. Finally, where AI has shone a spotlight on unregulated harms to workers that warrant broader interventions like issues related to data privacy or invasive workplace surveillance, Congress should advance new employment standards.

Pillar three: Increase workers’ ability to leave employers who create exploitative conditions with AI (or other technologies and practices) and ease job transitions

In earlier sections, we discuss how aggregate job loss is unlikely to occur as a result of advancements in AI technologies. For each job displaced by AI, there is highly likely to be one created by the ripple effects it creates (as has happened with other technological changes).  However, job churn creates stress and income losses for workers in the sector seeing job displacement from AI (Bivens and Zipperer 2024). If Congress is serious about easing the impact of AI-related job transitions for workers, then the most effective solution is strengthening social insurance programs to reduce the economic consequences of losing any specific job and provide adequate support to workers while they search for employment. 

In particular, Congress should strengthen unemployment insurance protections and take steps to reduce the cost of health care and increase access to retirement security. These, along with other improvements to social insurance systems and workforce development programs, will not only ease the financial burden and pain of specific job transitions, but increase the bargaining power of all workers—especially those who aren’t represented by a union. The single greatest bargaining chip nonunion workers have is the ability to leave their jobs. Therefore, efforts that improve our social insurance systems and reduce workers’ reliance on employers for health care and retirement, among other benefits, will increase their ability to walk away from subpar working conditions, creating pressure on businesses to change their practices. 

Similarly, it is far better to lose one’s job when overall unemployment is very low than when it is high. Further, threats to leave an exploitative workplace and find a better job elsewhere are far more credible when the aggregate labor market is experiencing very low unemployment than when unemployment is high. Macroeconomic policy that targets sustained periods of very low unemployment and that quickly restores the labor market to health after recessionary shocks is vital for workers to carve out a decent career during periods of technological changes. Of all the policy failures that created the rise in inequality and the anemic wage growth in recent decades, macroeconomic policy failures likely top the list. Sound macroeconomic policy will dwarf any level of AI penetration in its importance to the trajectory of workers’ wages and employment over the next few decades.

Conclusion

Over the last few years, researchers, unions, and workers have drawn much needed attention to how the deployment of AI technologies has exacerbated harmful workplace conditions and outcomes for many workers. It’s understandable why policymakers are focused on developing legislation that addresses the tool—AI technology—used by businesses to achieve these outcomes. However, we believe that legislation aimed at improving outcomes for workers must also address the underlying conditions that enable businesses to deploy AI in this manner. Legislative interventions must carefully consider and address why workers don’t currently have the bargaining power to directly engage employers when they use AI—or any technology—in a manner that skirts their rights or creates poor working conditions. That is why we believe any serious worker-centered AI policy must restore core labor rights, strengthen social insurance programs, and maintain full employment. If these underlying economic conditions are not addressed, then we fear that legislative interventions that regulate how AI is developed or deployed will fall far short of their objectives.

In the coming months, we will release a series of follow-up pieces that will examine each pillar, laying out in more detail how Congress and states could design specific proposals to achieve the policy objectives of each pillar. We believe the three broad policy recommendations discussed above would most directly and effectively increase workers’ ability to meaningfully engage employers over the deployment of AI technologies in the workplace. The advancement of technology creates an opportunity for greater prosperity and economic growth. Under the right conditions, workers can and should fully share in those gains.

But we will sound a note of caution: It is possible for policymakers looking to help workers thrive in labor markets to focus too much attention on AI-specific issues. Simply put, there is no compelling evidence that AI is enormously different enough or more powerful than other earlier waves of technological change in its effect on labor markets. Further, the pre-AI status quo in labor markets was terrible for workers’ opportunities to thrive and for needed fundamental policy change across all sorts of policy areas that remain unaddressed. Given this fact, U.S. workers would not be well served by having policymakers, researchers, and advocates focus disproportionate amounts of attention on AI deployment at the expense of other crucially needed reforms. To put it more bluntly, it would serve the interests of exploitative employers to have the pro-worker policy community lose focus on many other issues to concentrate large amounts of attention and influence on AI specifically.

Appendix

The summary below aims to distill federal bills related to AI by highlighting the most common policy interventions and strategies. By focusing on specific types of interventions, we hope to highlight both the nuance of the policy strategies as well as how even the same form of intervention can be structured and implemented in a variety of ways.

Disclosure requirements

Across branches and levels of government, there are efforts to increase basic awareness among workers (and the public) of the presence of AI technologies used in the workplace.

The Stop Spying Bosses Act broadly requires employers to disclose—to employees and the public—their practices related to workplace surveillance and how these affect or influence employment-related decisions. More specifically, the legislation requires employers to disclose the following: what data are collected, how the data are being collected, where and when data are being collected, how frequently the collection occurs, what the business purpose of the collection is, and which, if any, third party service providers are engaged for the surveillance, data transfer, or sale of worker-level data, among other disclosures.6

Similarly, the No Robot Bosses Act generally addresses employment-related decisions that are informed by outputs from automated decision systems. The bill would require employers to provide employees with a plain-language explanation of outputs from automated decision systems including: a description of the system used, description and copy of the data inputs to the system, an explanation of how the outputs were used in making the employment-related decision, and the reason for using the automated decision system outputs in making the employment-related decision.7

The Algorithmic Accountability Act directs the Federal Trade Commission (FTC) to implement, through regulations, a new requirement for companies to assess and disclose the impact of automated decision systems, including those related to “employment, workers management, or self-employment.” Specifically, covered entities would be required to submit an initial summary report of their assessment to the FTC prior to the deployment of newly covered technologies; entities would also be required to submit annual reports on the ongoing assessment of technologies that are already on the market. The bill identities specific topics that should be disclosed in the summary reports to the FTC including the following: the purpose of the product and a detailed description of the decision(s) the system intends to make; any publicly stated guardrails or limitations on use of the product; documentation of the data used as inputs during the development, testing, and maintenance of the systems; and any transparency mechanisms that are included in the system to allow end users to contest, correct, or appeal decisions made by the system.8 

The Warehouse Worker Protection Act broadly aims to address productivity quotas that are opaque to employees, increase risk of injury, or are used as pretext for exercising labor rights; the bill includes a number of disclosure requirements. Specifically, the bill would require covered employers to provide a written description of each quota an employee is subject to. The description must be provided in plain language and include the following: the number of tasks that must be completed within a specified time period; what, if any, disciplinary actions could result from failure to meet quotas; how the employer measures work speed, including where and when measurements occur; and the businesses’ purpose for collecting work speed data. Employers that take adverse employment actions against workers as result of failing to meet established quotas would be required to provide employees with a written explanation of how the quota wasn’t met and a copy of work speed data.9 

Auditing and impact assessment requirements

As discussed above, one of the key labor market impacts of AI is the potential for technologies to undermine existing labor and employment rights, including anti-discrimination protections. A number of policy proposals and state-level laws attempts to address these issues by creating new pre-deployment and ongoing auditing requirements of technologies, particularly those used in hiring or other employment-related decisions. 

Under the No Robot Bosses Act, employers would not be able to use any automated decision system to inform employment-related decisions unless it underwent pre-deployment testing and validation. Specifically the bill requires validation with respect to compliance with employment laws including: Title VII, ADEA, ADA, FLSA, Rehabilitation Act, and Pregnant Workers Fairness Act, among other specifically identified protected classes. The proposed legislation would also require, at minimum, annual independent auditing to ensure continued compliance with these laws.10

Under the Algorithmic Accountability Act of 2023, companies that develop AI technologies would be required to conduct pre-deployment and ongoing impact assessment for automated decision systems. The bill identifies an extensive list of issues that developers should include in their assessments including: comparison of performance outcomes under test conditions and deployment conditions; evaluation of differential outcomes associated with race, color, sex, gender, age, disability, religion, family status, socioeconomic status, or veteran status; assessment of the need for guardrails or limitations on the use of products; and assessment of the explainability and transparency of systems for consumers.11

Prohibition of specific technologies or use cases 

Beyond increasing transparency through disclosure requirements and imposing new auditing requirements, a number of legislative proposals prohibits specific uses of AI technologies. In some instances, these prohibitions simply reiterate that certain actions are not allowed under existing laws; in other instances, the prohibitions attempt to curtail the use of AI technologies in ways that may be harmful to workers given their limited bargaining power and ability to alter employer practices on their own. 

Under the No Robot Bosses Act, employers would not be able to rely exclusively on automated decision systems to make hiring, firing, disciplinary, or leave-related decisions, though they could still use the systems as an input or factor in a decision-making process. This type of prohibition attempts to remove perhaps the most extreme use cases of the technology as it relates to employment decisions. Importantly, though, it is currently quite common for business groups and companies to claim that technology alone doesn’t make these decisions.12

Under the Stop Spying Bosses Act, employers would be prohibited from using surveillance technologies when covered workers are off duty or in sensitive areas such as a locker room or restroom. Additionally, the bill would prohibit the use of these technologies to identify individuals who are engaging in labor organizing activities— a use of technology which the NLRB general counsel has noted is, in her office’s view, unlawful under the NLRA. Additionally, the bill would prohibit employers from using surveillance data to identify workers’ political opinions, religious views, and health conditions and outcomes that are unrelated to the performance of job duties.13

The Warehouse Worker Protection Act includes provisions that would prohibit employers from requiring a quota that would prevent the following: compliance with any required meal or rest breaks; compliance with health and safety standards (required by federal, state, or local laws); employees’ use of bathroom facilities; and compliance with reasonable accommodations and nondiscrimination provisions of federal, state, and local laws. Additionally, the bill would prohibit employers from setting quotas that establish a performance target over a period that is less than a day and prohibits quotas that would prevent or discourage employees from exercising their rights under the National Labor Relations Act.14

Amending existing labor and employment laws; agency directives 

As discussed above, businesses’ use of AI technologies has exacerbated violations of existing labor and employment standards. Some legislation recognizes this by explicitly prohibiting technology to be used in a manner that violates these rights, such as nondiscrimination protections or labor organizing rights (see above). In addition to this intervention, some proposed legislation goes further to amend underlying statutes to bolster and expand protections, and direct agencies to issue regulations and reports. 

The Warehouse Worker Protection Act would notably amend the National Labor Relations Act to make it an unfair labor practice for employers to “impose on an employee a quota that significantly discourages or prevents, or is intended to significantly discourage or prevent, an employee from exercising the rights guaranteed in section 7.” Additionally, the bill would amend the NLRA to create a rebuttable presumption of discrimination if an employer takes an action to impose a quota against an employee within 90 days of exercising their rights under section 7 of the NLRA. Additionally, the bill includes a series of directives to OSHA regarding rulemaking and amends the OSH Act. Specifically, the bill would require OSHA to issue the following: a proposed ergonomics standard (within 3 years) and a proposed standard requiring covered employers to provide employees access to trained first aid professionals at the facility (within 1 year). Finally, the bill would amend the OSH Act to include provisions related to the correction of serious, willful, or repeated violations pending contest and procedures for a stay. 15

The Eliminating Bias in Algorithmic Systems Act generally requires every federal agency to establish an office of civil rights that is focused on combating AI bias and discrimination; the bill is not limited to agencies that regulate employment practices but is inclusive of them. While the bill does not direct agencies to issue any specific piece of regulation or even amend their underlying authorities, the bill directs agencies to engage in specific activities in order to better identify and reduce prevalence of bias and discrimination. Specifically, the bill requires agencies to submit a report to Congress on the state of the technology with respect to the jurisdiction of the agency and any relevant steps the agency has taken to mitigate algorithmic bias and discrimination.16

Workforce and training

As discussed above, one concern among policymakers is the impact of AI on job displacement and skills. There are a few bills that address this issue, though primarily by directing agencies to engage in further research. 

The Technology Workforce Framework Act generally expands the functions of NIST to include a workforce framework for emerging technologies. Specifically the bill would direct NIST to define AI-related jobs and the necessary knowledge, skills, and abilities needed to fill those jobs.17

The Jobs of the Future Act of 2023 directs the Department of Labor and the National Science Foundation to submit a report to Congress on AI and its impact on the workforce including: industries that are projected to have the most growth in the use of AI technologies and whether that use would result in job enhancement or job replacement; analysis of the skill necessary to develop or use AI technologies; methods to ensure necessary skills, expertise, and education are accessible to all segments of the current and future workforce; recommendations to minimize job displacement; and workforce training needs.18

Notes

1. See for examples: Ellingrud et al. 2023; Goldberg 2023; Kochhar 2023; Tamayo et al. 2023.

2. See for examples: Bernhardt, Suleiman, and Kresge 2021; Yang 2020.

3. See for examples: Khan 2024; Long 2022; Mims 2021.

4. See BLS 2023. In 2019 total nonfatal injury and illness rates per 100 full-time workers was 4.8; in 2022 (latest data available), the rate was 5.5.

5. Richard L. Trumka Protecting the Right to Organize Act of 2023, [H.R.20] 118th Cong. (2023) and Public Service Freedom to Negotiate Act of 2024, [S. 4363] 118th Cong. (2024).

6. Stop Spying Bosses Act, [S.262] 118th Cong. (2023).

7. No Robot Bosses Act, [S.2419] 118th Cong. (2023).

8. Algorithmic Accountability Act of 2023, [H.R.5628] 118th Cong. (2023).

9. Warehouse Worker Protection Act, [S.4260] 118th Cong. (2024).

10. No Robot Bosses Act, [S.2419] 118th Cong. (2023).

11. Algorithmic Accountability Act of 2023, [H.R.5628] 118th Cong. (2023).

12. No Robot Bosses Act, [S.2419] 118th Cong. (2023).

13. Stop Spying Bosses Act, [S.262] 118th Cong. (2023).

14. Warehouse Worker Protection Act, [S.4260] 118th Cong. (2024).

15. Warehouse Worker Protection Act, [S.4260] 118th Cong. (2024).

16. Eliminating Bias in Algorithmic Systems Act of 2023, [S.3478] 118th Cong. (2023).

17. Technology Workforce Framework Act of 2024 [S.3792] 118th Cong. (2024).

18. Jobs of the Future Act of 2023, [H.R.4498] 118th Cong. (2024).

References

Athena Coalition, National Employment Law Project, Warehouse Workers for Justice, Warehouse Worker Resource Center, United for Respect, Atwood Center, Make the Road New Jersey, Make the Road New York, and New York Communities for Change. 2024. Packaging Pain: Workplace Injuries in Amazon’s Empire, January 2020.

Bernhardt, Annette, Reem Suleiman, and Lisa Kresge. 2021. Data and Algorithms at Work: The Case for Worker Technology Rights. UC Berkeley Labor Center, November 2021.

Bivens, Josh, and Ben Zipperer. 2024. Unbalanced Labor Market Power Is What Makes Technology—Including AI—Threatening to Workers, Economic Policy Institute, March 2024.

Bureau of Labor Statistics (BLS). 2023. “Table 1. Incidence Rates of Nonfatal Occupational Injuries and Illnesses by Industry and Case Types, 2022,” Injuries, Illnesses, and Fatalities, 2022. Last modified November 8, 2023.

Center for Labor and a Just Economy Lab (CLJE Lab). 2024a. Building Worker Power in Cities and States: A Toolkit for State and Local Labor Policy Innovation. Harvard Law School, September 2024.

Center for Labor and a Just Economy Lab (CLJE Lab). 2024b. Worker Power and Voice in the AI Response. Harvard Law School, January 2024.

Chapman, Michelle, and Haleluya Hadero. 2024. “Amazon Labor Union Members Vote Overwhelmingly in Favor of an Affiliation with the Teamsters.” Associated Press, June 18, 2024.

Clean Slate for Worker Power (Clean Slate). 2021. “Principles of Sectoral Bargaining: A Reference Guide for Designing Federal, State, and Local Laws in the U.S.” Harvard Law School’s Labor and Worklife Program, May 2021.

Department of Labor (DOL). 2024. Memorandum, Subject: Artificial Intelligence and Automated Systems in the Workplace under the Fair Labor Standards Act and Other Federal Labor Standards. April 29, 2024.

Doellgast, Virginia, Sean O’Brady, Jeonghun Kim, Della Walters, Alyssa Acevedo, Nelson Dragsbaek, Tom Hegeman, Stefan Ivanovski, Natalie McCormick, Deepa Rajan, Ayaj Rana, MT Snyder, Jelena Starcevic, and Ines Wagner. 2023. AI in Contact Centers: Artificial Intelligence and Algorithmic Management in Frontline Service Workplaces. Cornell ILR School, November 2023.

Ellingrud, Kweilin, Saurabh Sanghvi, Gurneet Singh Dandona, Anu Madgavkar, Michael Chui, Olivia White, and Paige Hasebe. 2023. Generative AI and the Future of Work in America. McKinsey Global Institute, July 2023.

Equal Employment Opportunity Commission (EEOC). 2023. ” EEOC Releases New Resource on Artificial Intelligence and Title VII” (press release), May 18, 2023. 

Glass, Aurelia. “Unions Give Workers A Voice Over How AI Affects Their Jobs” Center for American Progress. May 2024.

Goldberg, Emma. 2023. “A.I.’s Threat to Jobs Prompts Question of Who Protects Workers.” New York Times, May 23, 2023.

Griffith, Kati L. 2019. “The Fair Labor Standards Act at 80: Everything Old Is New Again.” Cornell Law Review 104, no.557 (2019). https://scholarship.law.cornell.edu/clr/vol104/iss3/1.

Khan, Mishal. 2024. At Work and Under Watch: Surveillance and Suffering at Amazon and Walmart Warehouses. Oxfam, April 10, 2024.

Kochhar, Rakesh. 2023. Which U.S. Workers Are More Exposed to AI on Their Jobs? Pew Research Center, July 2023.

Lang Gordon, Julia. 2021. “Under Pressure: Addressing Warehouse Productivity Quotas and the Rise in Workplace Injuries.” Fordham Urban Law Journal, 2021 49, no. 1. https://ir.lawnet.fordham.edu/ulj/vol49/iss1/5.

Long, Katherine. “Amazon Workers Say Minor Aches Suddenly Became Debilitating as They Raced to Meet Speed Targets.” Business Insider, October 19, 2022.

Madland, David. 2019. How to Promote Sectoral Bargaining in the United States. Center for American Progress, July 2019.

Mims, Christopher. 2021. “The Way Amazon Uses Tech to Squeeze Performance Out of Workers Deserves Its Own Name: Bezosism.” Wall Street Journal, September 11, 2021.

National Labor Relations Board. 2022. “NLRB General Counsel Issues Memo on Unlawful Electronic Surveillance and Automated Management Practices” (news release), October 31, 2022.

National Nurses United (NNU). 2024. “National Nurses United Survey Finds A.I. Technology Degrades and Undermines Patient Safety” (press release), May 15, 2024.

O’Brady, Sean, and Virginia Doellgast. 2021. “Collective Voice and Worker Well-Being: Union Influence on Performance Monitoring and Emotional Exhaustion in Call Centers.” Industrial Relations: A Journal of Economy and Society 60, no. 3 (July): 307–337. https://doi.org/10.1111/irel.12286

Occupational Health and Safety Administration (OSHA). 2023a. “Department of Labor Announces National Emphasis Program Aimed at Reducing, Preventing Workplace Hazards in Warehouses, Distribution Centers” (news release). U.S. Department of Labor, July 13, 2023.

Occupational Health and Safety Administration (OSHA). 2023b. “OSHA nstruction: National Emphasis Program on Warehousing and Distribution Center Operations.” U.S. Department of Labor, July 13, 2023.

Occupational Safety and Health Administration (OSHA). 2024. “Ergonomics – Standards and Enforcement FAQs” (web page). U.S. Department of Labor. Accessed on September 4, 2024.

Rosenberg, Eli. “Fired Amazon Union Organizer at Kentucky Warehouse Alleges Retaliation.” NBC News August 25, 2022

Salaam, Ameenah. 2023. “Written Comments for AI Insight Forum on Workforce.” Washington, D.C., October 27, 2023.

Tamayo, Jorge, Leila Doumi, Sagar Goel, Orsolya Kovács-Ondrejkovic, and Raffaella Sadun. 2023. “Reskilling in the Age of AI.” Harvard Business Review, September 2023.

Tung, Irene, Nicole Marquez, and Paul K. Sonn. Amazon’s Outsized Role: The Injury Crisis in U.S. Warehouses and a Policy Roadmap to Protect Workers. National Employment Law Project, May 2024.

Yang, Jenny R. 2020. Three Ways AI Can Discriminate in Hiring and Three Ways Forward. Urban Institute, February 2020.