This week, the conservative think tank American Compass released some thoughts about the impact of AI on work titled “For Whom the Machine Toils”. Their collection includes this essay, which I have lightly adapted.
Nick was the shop steward from Central Casting. A chain-smoking Greek immigrant and World War II veteran, he welcomed me to my new job at the largest machine shop west of the Mississippi. Any time he saw me working too eagerly, I would feel his grandfatherly hand on my shoulder as he advised me to, “Stretch things out. Make the work last.”
As a machine operator, my job was to load small steel parts into a mill that turned them into military-grade turbine blades. When I began helping engineers automate this tedious work, Nick warned me about “killing good union jobs.” He distrusted automation, especially the computerized machine tools that were transforming the massive Westinghouse defense plant in the heart of Silicon Valley.
I liked technology and took night school classes to learn to program the new machines. Eventually, I found a way to placate Nick and his fellow technophobes by reminding them that “machines that make more can pay more.” Whether higher pay would come from our solidarity, our employer’s need for specialized skills, or a buoyant labor market that rewarded higher productivity did not concern us. We trusted each other. Our union made us strong.
The Stakes for Labor
Like Nick, frontline workers today have good reason to feel ambivalent about coming workplace technologies. They can rarely influence how new tools are designed or deployed. They have little visibility into the gains that result, nor any say in how they are shared. They are often unsure what new skills they need or how best to acquire them.
The enormous scale of US investment in artificial intelligence technologies increases risk. Currently, AI infrastructure spending fuels most of the US economy's growth. It also drives the majority of stock market and startup expansion. Researchers at the Federal Reserve estimate that companies are adopting AI technologies at twice the rate they adopted the internet.
This represents a huge gamble, especially in a business with modest profitability. The gold rush to invest in massive AI data centers could represent a bubble that, when it bursts, causes millions of working families to suffer. US companies plan to invest about $500 billion in AI infrastructure, an amount that seems to grow weekly. The bulk of this money will go to specialized computer chips (GPUs), which lose value much more quickly than assets like railroads, buildings, or telecommunications lines that companies overbuilt during past investment booms. The risk of a crash increases, especially when companies take measures to conceal their massive loans to preserve the appearance of creditworthiness. Throughout American history, funding manic investments in a single asset has been the disease, not the cure.1
Of course, it is the perennial fear of AI-driven job losses that captures daily media attention. The CEO of a leading AI company recently warned that “due to AI job losses, ordinary people will lose their economic leverage, which breaks democracy and leads to severe concentration of power. This technology will be a disaster not only for American workers, but for American democracy.”
These forecasts may be correct, but even highly qualified people often overhype new technologies and overestimate their impact on jobs.2 Americans typically face a greater risk that technology will improve productivity too little, rather than too much. Of course, this time could be different. AI may not be just another round of automation. Its reach into non-routine cognitive work, its speed of improvement, and the low barriers to global deployment make it different enough that we need to ask with new urgency how workers can have a say in AI’s deployment and share in its gains.
Traditional labor unions cannot offer much help. Not only are unions nearly extinct in the private sector, but unions face difficulty in aligning technology with the interests of their members. This is partly because most workplace technologies are profoundly unpredictable. It is also because new technology typically affects every worker in a sector, but our labor laws permit unions to negotiate, at best, with one company at a time.
Absent a legal framework to support sector-wide agreements, unionized companies fear they may end up at a competitive disadvantage and are reluctant to bargain over new technologies. Faced with diminishing influence, technological uncertainty, and perverse incentives, many unions become reflexively hostile to new technology.
As the AI tsunami approaches, are there steps we can take to amplify worker voices short of redesigning labor law from the ground up? Can we structure new forms of worker power to compel capital to invest productively and ensure that workers receive a fair share of any returns that result?
Technological Uncertainty
Nick worried about computerized machine tools in part because he could see old jobs leaving much more clearly than he could see new ones arriving. He was hardly alone: there are several reasons that humans struggle to predict the impact of technological change.
Complements or substitutes? To start, we cannot always know which tools enhance human skills and which replace them. The same technology that simplifies a task today may help eliminate it tomorrow. At first, the internet empowered travel agents to discover new destinations. Later, it enabled consumers to book their own travel directly, resulting in the disappearance of 70% of travel agent jobs over the next two decades. More recently, travel agent jobs have recovered, in part thanks to the creative use of AI tools.3 At first, ATMs enhanced the work of bank tellers. Today, mobile banking, not ATMs, is making tellers obsolete.4 These patterns—initial complementarity, followed by obsolescence—make technological forecasting a treacherous endeavor.
Structure of demand. Whether costs lowered by technology reduce jobs depends on the structure of demand. When demand is highly elastic, sales and employment increase as costs fall. As technology and scale reduced the cost of solar power, demand exploded, creating thousands of new jobs. In contrast, as the cost of growing food fell, demand barely increased, so we needed fewer farmers. These effects can be challenging to gauge in advance.
Tasks vs jobs. Automation eliminates more tasks than jobs. Current research suggests that LLMs will affect at least 10% of the tasks performed by 80% of US workers. Twenty percent may see most of their tasks affected. However, the number of jobs created or lost remains unknowable, in part because the same technology that eliminates old tasks enables new ones that were previously impractical.
As chatbots handle repetitive inquiries, such as password resets and order tracking in multiple languages, human customer service agents will focus on more complex, sensitive, or high-value cases where human judgment is most crucial. Or they may focus on new service offerings, such as helping small businesses expand globally with low-cost, multilingual support. Contrary to the predictions of AI godfather Geoffrey Hinton, AI tools for reading medical scans have not eliminated radiology jobs. So far, they appear to complement human skills, enabling doctors to spend more time with patients.5 It is very challenging to anticipate the balance between old tasks that technology has destroyed and new ones that it has enabled.
Impact on culture. Technology can affect not only jobs, skills, and pay—it can also transform culture in ways we cannot foresee. Railroads not only promoted the growth of new industries but also transformed the publishing of books, fashion, and newspapers. Agricultural mechanization accelerated urbanization and the growth of the postwar consumer economy. Nobody predicted these second- and third-order results when the railroad came to town or machines came to the fields. Likewise, nobody can predict the full cultural consequences of social media reducing long-form reading, LLMs killing off essay writing, or children deciding that their best friends are anthropomorphic chatbots. Knowing that technology can drive significant changes in how our children grow up induces as much uncertainty as worries about the impact on our jobs.
One purpose of an economy is to support stable livelihoods, families, and communities. Everyday workers want to know what is coming around the corner, especially if it affects their paycheck or way of life. They recognize that decisions about how companies implement new technologies have significant implications for their work, family, and community. Still, they have very few ways to formulate or express their preferences. Without a voice, technological uncertainty gives rise to political anxiety.
Amplifying Worker Voices
Especially when their agreements cover the leading companies in a sector, some unions have tried to influence how industries design and deploy new technologies. The United Auto Workers required automakers to retrain displaced workers, limit layoffs in certain plants, and share productivity gains through profit sharing. The Hollywood writers and actors strike tried to restrict studios from using AI to write or rewrite scripts without human credit, or to digitally reproduce an actor’s likeness without consent and compensation.
These agreements are, at best, minor victories. Unions exert limited influence on workplace technology, for the same reason they exert limited influence on other matters: the United States has very few private sector unions, especially in low-pay, low-productivity sectors. Private sector union density in the United States peaked at about 35% in 1954. Unions now represent fewer than 6% of private sector workers, many of whom work in protected sectors.6 Without significant legal reforms, America’s current model of collective bargaining cannot give most American workers a say in how companies deploy new technologies or share the benefits that result.
Many proposed alternatives for amplifying worker voices—such as company-run forums, surveys, or ad hoc committees—don’t give workers an independent, durable structure through which to organize and think collectively. Workers need more than a channel to “speak up.” They need organizations of their own—whether unions, professional associations, or guilds—that allow them to consult external experts, analyze complex data, debate trade-offs, and form a collective view that management and policymakers will take seriously.
This is not possible if workers are left to respond as isolated individuals. Faced with uncertainty about how new technologies might reshape jobs, their instinct is often to resist change rather than weigh the potential benefits and risks.
Even without stronger unions, however, labor activists, managers, communities, and public officials can advocate for mechanisms to give front-line workers greater influence over how companies use AI technologies.
American co-determination. The case for worker representatives on boards of directors is partly democratic. Workplaces are central to our lives, and board representation gives those affected by decisions a voice in making them. There is also a practical case. I served as a union representative on the board of a steel company that emerged from a financial restructuring with substantial employee ownership. Workers benefited from having a trusted representative distill information about the company’s performance and market challenges. Board discussions were less likely to pursue short-term strategies that overlooked workforce realities. Board representation gave workers a voice in decisions that shape their wages, benefits, job security, and new investments.
Germany and most Scandinavian countries have done this for many years. Peer-reviewed estimates suggest that these efforts do not raise pay, output, or layoffs, but do increase business investment. Nor is the idea especially partisan. Senate liberals have championed worker representation on corporate boards, as have conservative voices, including Chris Griswold at American Compass and Senator Bernie Moreno (R-OH).7
Works councils. Implementing any new technology, especially one as potentially transformative as LLM-based tools, benefits from improved information flow and joint problem-solving. One tested approach is the use of elected shop-floor committees.
Elected works councils provide employees with a formal, elected channel to raise concerns, share insights, and influence workplace decisions without relying on management-controlled structures. Councils address concrete workplace issues—such as scheduling, health and safety, training, and technology adoption—enabling quicker and more tailored solutions. They provide a regular forum for discussion that helps reduce conflict and build trust between workers and employers. Works councils can mediate transitions—such as restructuring, automation, or new processes—by giving workers input into how companies implement changes, making technology adoption smoother.
Pro-worker conservatives in Congress who wish to promote works councils could start by excluding them from the National Labor Relations Act (NLRA) Section 8(a)(2) ban on worker organizations receiving employer funding, while protecting against the re-emergence of company unions.8
Wage boards. Wage boards like New York’s Fast-Food Wage Board assemble representatives of workers, employers, and the state to regulate pay, hours, training, and skills certifications. These can be a way for worker voices to influence the introduction of AI tools in the workplace. By setting consistent labor and skill standards across all firms in a sector, wage boards deter companies from seeking a low-cost advantage and encourage managers to compete on operational efficiency. Studies have found that these boards can increase labor productivity while reducing churn, particularly in low-margin service sectors.
Joint apprenticeship programs. Multi-employer partnerships—often involving a community college and a worker trustee—already run paid apprenticeships and issue nationally portable credentials. In the Department of Labor’s 2022 American Apprenticeship Initiative, 99% of employers reported indirect benefits, and 84% said those benefits were at least as significant as the direct productivity bump from the apprentice’s output. Because apprenticeships embed remedial instruction alongside on-the-job training, they are an effective way to enhance the skills and lifetime earnings of workers who would otherwise stagnate.
Congress should significantly expand these programs with subsidies to community colleges and employers. One promising employer-subsidy model is the American Workforce Act, introduced in the 118th Congress by Senator Tom Cotton (R-AR), which would establish a workforce training grant program that would provide companies with up to $9,000 per worker for earn-as-you-learn on-the-job training.9
Worker centers. In the face of NLRA restrictions on labor unions in the United States, worker centers have proliferated. These community-based institutions provide a range of services and support to low-wage workers. They are not unions, but typically focus on issues such as wage theft, unsafe working conditions, and inadequate access to resources. They have set innovative standards for some workers and could play a valuable role in helping to set AI standards. A report from MIT counted at least 246 active worker centers at the end of 2021.10
For example, the Coalition of Immokalee Workers has pressured major fast food and grocery chains to pay a small premium on produce to support farmworkers and to purchase only from farms that have signed the Fair Food Code of Conduct, which guarantees rights such as access to water and shade. Likewise, the Restaurant Opportunities Centers, United for Respect (an amalgamation of smaller worker centers), and the National Domestic Workers Alliance combine legal enforcement, skills boot camps, and public advocacy.
High-involvement work systems. Companies like Costco and Toyota have long recognized the value of lean teams, cross-training, stable work schedules, and incorporating worker voice into continuous improvement schemes to reduce waste and streamline operations. Unsurprisingly, companies that adopt these models experience substantial reductions in turnover (approximately 25%) and significant productivity increases (approximately 15%) within two years.
Historically, unions have been hostile to these initiatives because they can seem to replace unions as forums for worker representation. As a leader in the Machinists Union, I developed a dozen unreasonable demands designed to subvert workplace participation schemes that threatened to supplant union shop-floor primacy. Eventually, a savvy aerospace executive called my bluff by simply agreeing to all of them.
Unions should continue to invest in an ecosystem of pro-worker organizations to promote initiatives like these. Unions representing private sector workers take in roughly $5 billion in dues each year.11 They should test whether investments in strengthening pro-worker organizations can reduce the cost and expand the reach of NLRA-style labor organizing, which produces only 40-90,000 new members each year. (This rounds to zero in a private-sector labor market of 135 million workers that adds and removes 25-30 million jobs each year.)
Investing in experiments and alliances could seed a network of reinforcing initiatives that can adapt quickly as work, workers, and workplaces evolve. Most of the initiatives outlined above fall outside America’s historic model of collective bargaining; yet, they provide workers with levers to influence pay, schedules, and workflow design.
Conclusion
Most companies adopt new technologies through a series of choices made by managers and vendors, which give workers little say in how the tools are designed or deployed. Without organizations that can help them understand, influence, and benefit from those choices, we will once again ask workers to navigate profound disruptions on their own. If instead, we can reimagine worker power—through unions designed for current conditions, worker board representation, training partnerships, or new forms of digital and community-based organizing and bargaining—then workers may be able to harness technology to serve their interests, not just those of business owners.
In time, I came to appreciate Nick’s skepticism of new technology. He understood that in the absence of trust and shared power, workplace technologies can lubricate exploitation. In our shop, a well-organized union prevented that. It gave us a way to bargain not just with our employer, but to understand each other better. For AI to deliver not just workplace innovation but also dignity, we need that kind of solidarity in every American workplace.
Coda
Massive speculation in railroads led directly to economic panics and widespread misery in 1857, 1873, and 1893. Recessions rarely have a single cause, but asset bubbles resulting from overbuilding are generally thought to have contributed to depressions in 1920 and 1929, the savings & loan crisis of the 1980s, the dot-com and telecom bust of 2000, and the Great Recession of 2007.
Herbert Simon was a deeply respected AI pioneer and remains the only person to win all three of the most coveted awards in US computer science: the Turing Award, the Nobel Prize, and the National Medal of Science. In a famous 1965 book, Simon confidently predicted that, “Machines will be capable, within twenty years, of doing any work a man can do.”
Since the pandemic, travel agent jobs have recovered significantly as travelers seeking complex adventures, cruises, and personalized experiences seek out professional expertise. Some analysts view AI tools as a means to enhance the productivity of travel agents. Others argue that AI will once again empower users and disintermediate agents.
Automatic teller machines famously led banks to open more branches, so although each branch needed fewer tellers, the number of bank tellers in the United States doubled from about 300,000 in 1970 to around 600,000 by 2010. Additionally, as routine cash handling became automated, tellers shifted toward providing higher-value personalized service, sales, and more complex customer interactions. Then digital banking reversed this trend. Apps have made in-branch visits less necessary, so banks need fewer branches and tellers. The Bureau of Labor Statistics reports that the number of full-time teller jobs dropped to approximately 364,100 by 2022. They project a decrease of about 8–12% in teller jobs by the late 2020s.
As the terrific Works in Progress post makes clear, there are many reasons that AI tools have not (yet) taken over for radiologists, including performance gaps between benchmarks and clinical performance, the fragmented and specialized nature of AI models, regulatory, legal, and institutional constraints, tasks performed by radiologists that extend beyond diagnosis, demand elasticities that have led to more scans being ordered, and adoption frictions that lead hospital adoption of advanced technologies to lag the technical frontier.
Unionized private sector workers are concentrated in legacy sectors or those with less competitive pressure. Utilities are regulated natural monopolies, and approximately 18% of their workforce is unionized. Commercial construction (where unions influence procurement decisions) is 18% unionized among craft workers. Transportation (where competition is restricted by expensive infrastructure and regulatory licenses) is 16% unionized. Private universities and other private educational services are 13% unionized. In contrast, finance, insurance, software, professional or technical services, agriculture, or hospitality are all less than 2% unionized.
Senator Elizabeth Warren (D–MA) and Senator Tammy Baldwin have co-sponsored legislation requiring corporations with over $1 billion in annual revenue to obtain a federal charter that emphasizes stakeholder responsibility. This bill also mandates that employees elect 40% of corporate board members. Senator Bernie Sanders (I–VT) has consistently advocated for worker seats on corporate boards.
In 2024, during the 118th Congress, then-Senators Marco Rubio and J.D. Vance sponsored legislation to do this.
The bill was co-sponsored by then-Senator J.D. Vance (R–OH), and a companion version was introduced in the House by Rep. Max Miller (R–OH). In addition to providing a $9,000 federal voucher for high school graduates or GED holders without a bachelor’s degree to enter employer-led workforce training programs, the bill offers a $1,000 bonus to employers for each trainee hired after completing the program. To qualify, positions must pay at least 80% of the local median household income. Employers can deliver training directly or through trusted partners (e.g., trade associations, community colleges, or labor unions). E-Verify is required for participating employers. The program would be funded partly by a 1% wealth tax on large private college endowments with more than 500 full-time students and endowment thresholds of over $2.5 billion or $500,000 per student, unless the college is religiously affiliated.
The rapid growth of worker centers in the United States dates back to the 1990s; however, these organizations have a much longer history. The Silicon Valley AFL-CIO helped establish one of the first in 1982 as the Silicon Valley Toxics Coalition. We sought to educate semiconductor assembly workers at large wafer fabs run by Intel, AMD, National Semi, and Fairchild. These immigrant women were exposed daily to TCE (trichloroethylene), a solvent that the EPA recently banned as a carcinogen.
The BLS estimates that the average unionized private-sector worker earns $65,000/year. Assuming union dues of 1.5% of wages, six million private sector union members generate $5.85 billion of dues revenue annually. This calculation excludes members of professional associations who do not engage in collective bargaining but includes represented workers in right-to-work states who do not pay dues.