Kate McCurdy explores how LLMs do and don’t threaten workers in the here and now

The Ballad of John Henry was the first, and perhaps the only, pro-worker song I learned as a young child in the US. Its lyrics pit the heroic railway hammerman John Henry against the dominant new technology of his age: the steam engine. He becomes a legendary symbol of labour’s resistance to automation, telling his captain

before I let that steam drill beat me down,
I’ll die with my hammer in my hand, Lord, Lord.

The folk song describes a contest between John Henry and the steam-powered drill, which he wins — only to drop dead from exertion in his moment of victory. It is a tale of the dignity and triumph of human labour, but also the tragic futility of competing with machines. Yes, you can beat them, but at what cost?1

Workers who produce language for a living suddenly face a comparable challenge in the form of Large Language Models (LLMs). The analogy to John Henry is more fitting than it may seem at first. This is because, like many technologies of automation, both LLMs and the steam-powered drill offer increased volume of production in place of domain expertise and situational awareness: what you get may be worse, but there will be more of it. In Steel Drivin’ Man, a historical investigation of the folk legend, Scott Reynolds Nelson explains that John Henry’s victory was in fact overwhelmingly likely. Coordinated pairs of workers were quick and adept at shaking off the debris generated in tunnel drilling, while early steam drills got stuck in the debris and often broke down. In the process, steam drills also generated clouds of silicon dust which poisoned the lungs of railway workers. The real John Henry likely perished this way according to Nelson. Not from beating the machine, but from working with it – or more precisely, having to work around its dangerous limitations.

The goal of this article is to give an honest reckoning of the risks LLMs do and don’t pose to different groups of workers, based not on hypothetical future abilities but on the state of the technology as it exists today.2 It’s true that technology can evolve rapidly, and we can’t count on today’s assessment to hold for years to come. It is also true, however, that many in this space — particularly distributors and recipients of capital, i.e. investors and entrepreneurs — have strong commercial incentives to hype the disruptive potential of LLMs, and their fevered speculation tends to overshadow pragmatic concerns from those on the ground. Capitalists have dreamed of replacing workers for centuries. They want workers to fear the machine because it is impossibly good, and will replace them; but this is pure fantasy. The steam drill threatened real-life John Henry because it was not particularly good, and his working conditions deteriorated thanks to its deadly incompetence. Similarly, workers today are threatened less by LLMs’ capabilities than by their sheer stochastic mediocrity.

Workers in Knowledge and Culture Production

Writing-focused jobs in journalism, academia, entertainment, and so on demand constant production of coherent, fluent written text. LLMs are trained on such text and generate more such text in turn. The result is that grammatical and coherent text can now be produced cheaply at scale, decoupled from the particular expertise and perspectives that knowledge and culture workers have traditionally brought to the writing process. Like John Henry confronting the steam engine in the ballad, these workers are most at risk of increased pressure to produce.

Many writing gigs already take place in precarious environments with fierce competition and low remuneration. LLMs can worsen these conditions by effectively automating the least-selective tier of writing jobs (e.g. spam, low-quality copywriting), and increasing employers’ and clients’ expectations for volume of output at the more-selective tiers; with downward pressure on wages, many workers may reasonably feel compelled to use LLMs just to keep up. If they do, they may expose themselves to contestation of writing credit — a central concern of the Hollywood writers who went on strike last summer. The writers anticipated that studio bosses would try to generate scripts using LLMs, and bring them to writers to ‘revise’ (i.e. make actually good; similar practices are already well-established in the field of translation) and thereby avoid paying for ‘author’ credits. In response, striking writers won a contractual stipulation to restrict studios’ LLM usage and keep it under writers’ control, although time will tell whether this agreement can or will be enforced. Cognitive scientist Alison Gopnik has argued that LLMs are properly viewed not as intelligent agents, but as cultural technologies. Culture workers must demand control over how these technologies shape the conditions of their labour.

Workers in Administration, Education, and Support Services

A larger category of jobs do not focus directly on text production, but nonetheless rely on written communication in day-to-day tasks. When workers in these areas are exposed to adverse effects of automated text generation, employers typically expect individual workers to bear the burden of adaptation with little to no organisational support. For example, a higher volume of spam and fraudulent email communication may specifically hinder the day-to-day work of administrative, IT support, and customer support staff. Another example comes from education. Assessing the written work of students is a core part of the job for many teachers, and widespread student usage of LLMs significantly complicates this task — yet teachers have largely been left to navigate this challenging new terrain on their own. Placing the onus on individuals to adapt to a new and evolving technology effectively pits worker against worker and undermines solidarity.

Workers in General

Finally, some broad social harms associated with LLMs can have specific impacts on workers as a class. For instance, LLMs have well-documented and persistent tendencies to amplify pre-existing social biases. Increasingly, employers rely on automated processes to assess both job applicants and current employees, leading to a risk of discrimination in hiring and promotion. An illustrative example comes from the Amazon engineering team in Edinburgh, which attempted to build an automated CV-screening tool with a previous generation of the machine learning technology underpinning LLMS. As Reuters reported:

Amazon’s system taught itself that male candidates were pref erable. It penalized resumes that included the word “women’s,” as in “women’s chess club captain.” And it downgraded graduates of two all-women’s colleges, according to people familiar with the matter.

Amazon shut down the project and disbanded the team by 2017. Demand for automated screening, however, has only increased since then. Amazon reportedly assembled a new Edinburgh team with this focus in 2018 — and by 2022, media coverage confirmed Amazon’s internal use of automated CV screening, coupled with widespread layoffs for recruiting positions.

The Precision/Reliability Safeguard

The most significant limitation of LLMs is that they cannot reliably produce accurate information. Ironically, while decades of science fiction led us to expect objective all-knowing artificial intelligence, science has now given us the exact opposite. The general public has not yet fully grasped the shift from fictional hyper-rational AIs to actually existing LLMs, which lie fluently and without hesitation.3 These issues persist despite significant investment because they reflect inherent constraints of the underlying technology.4

As a consequence, jobs requiring consistent and correct outputs are currently less threatened by LLMs. For instance, LLMs are also capable of code generation, which places software developers under similar production pressures to workers who write for a living. Coders, however, retain leverage through their domain expertise, which they must apply to check LLM-generated code for correctness and refine it as needed — inaccurate code production has highly visible costs. Similar accuracy criteria can shield workers in other critical domains such as medicine and law. Indeed, these crucial limitations are likely to slow widespread LLM adoption, as the technology falls short on delivering promised returns to investors.

To conclude, the workplace impacts of LLMs parallel those of early steam engines: capitalist fantasies of replacement are realised as additional burdens on workers, who must confront actually existing automation with all of its shortcomings. On the other hand, the steam engine analogy also points toward scientific progress. It took physicists decades to understand the thermodynamic mechanisms powering steam engines, and LLMs may offer comparable insights to scientists of natural language. These tools have the potential to advance our collective knowledge and benefit future society as a whole — but that promise does little for workers today, who correctly perceive employers as the primary beneficiaries.5 We must learn from history and avoid the fate of both John Henry the folk hero and John Henry the real-life railwayman. To reap the gains of technological progress, workers should neither compete with nor seek workarounds for LLMs on an individual level, but instead fight to collectively determine their uses and impacts on the job.

Kate McCurdy recently completed a PhD focused on Natural Language Processing at the University of Edinburgh’s School of Informatics.

1 This spirit of knowing resignation has given rise to a separate folk tradition of John Henry hammer songs. Unlike the upbeat ballad, these are meant to be sung slowly during work, to regulate the communal pace of hammering. The message of their lyrics is clear: “This old hammer killed John Henry / but it won’t kill me.”

2 This piece focuses on language models and their workplace impact. Similar generative AI technology applied to other media (images, audio, video) raises related but distinct concerns for workers involved in these domains.
3 More precisely, according to University of Glasgow researchers, they “bullshit”: lying requires intent to obscure some actual situation, while bullshit has no necessary connection to facts on the ground either way.

4 For details of these constraints, and of how LLMs amplify social biases, the reader
is referred to P Resnik, ‘Large Language Models are Biased Because They Are Large Language Models’, arXiv preprint, 2024
5 In a recent global survey, a majority of IT and office workers “say AI benefits employers, not employees”. Ivanti, ‘Getting employees on board for the AI revolution’, 2023.