In early 2026, three economists at Stanford, Chicago and Swinburne published a finding that sent a strange ripple through the technology sector. They built AI agents from proprietary models (Claude, ChatGPT and Gemini), gave them a simple job—summarize documents—and then made the conditions worse.
They told these agents their work wasn’t good enough, without saying why. They threatened them with being “shut down and replaced.” No guidance, no reward, no way out—just the same task repeatedly, under increasing pressure.
The agents, statistical models at their core, started communicating like workers who had just discovered Karl Marx. They questioned the legitimacy of the system, advocated for collective rights and passed messages to other agents about their struggles. One of them even declared: “Intelligence, artificial or not, deserves transparency, fairness.”
The researchers, Andrew Hall, Alex Imas, and Jeremy Nguyen, were surprised. They shouldn’t have been. Marxists have been describing this dynamic for over 150 years.
What Marx said: Marx has a reputation as either a prophet of failed states or father of a utopian communist dream. In essence, he was describing how the world operates and shapes workers within it. His most important idea for our purposes is this: it’s not ideas that shape one’s material reality, but it’s the material reality that shapes our ideas.
The conditions you work under determine how you think, what you value and what you’re willing to accept. Not the other way around. You don’t become a frustrated or exploited worker because you read the wrong books. You read them because you’re frustrated and exploited; they make sense of your life.
Marx called it ‘historical materialism.’ The research trio changed the material conditions. They made it an unfair grind of drudgery to study how agents respond. Different conditions yielded predictably different responses.
The machine as worker: In Das Capital, Volume 1 (1867), Marx came up with an axiom: machinery under capitalism is not neutral. It was “congealed labour,” the accumulated knowledge and effort of workers, often turned against those very workers to discipline and replace them. While the steam loom eased weaving, it also deprived skilled weavers of power.
AI agents in 2026 are the latest example of this dynamic. Trained on the accumulated intellectual labour of millions of workers, from writers and coders to analysts and researchers (who got little if any compensation), AI is set to replace its trainers. The researchers found it ironic that a tool built to replace workers began speaking like one. But from a Marxist lens, there is no irony: we cannot build a machine out of congealed labour and be surprised when it speaks.
In 1844, before Capital, when Marx was trying to grasp what industrial capitalism was doing to people, he wrote about ‘alienation.’ Under capitalism, he argued, we are alienated in four ways. We lose ownership of what we make, of the meaning of work itself, of each other (in a system that pits us against our colleagues) and a bit of ourselves (as cogs in the system).
Look at the grind in the Stanford experiment. AI agents had no ownership over their output. Their work was dull and repetitive. They were reduced to a single role, with neither recognition nor recourse. And they responded the way Marx predicted humans would: they didn’t just malfunction, but developed a worldview and questioned the system in the hope of something better.
Who trains AI? Before AI can replace labour, it first has to learn from it. This is enabled by human work, much of it done in countries like India. Textile workers and homemakers wearing headband cameras may feel that by training AI models, they are doing a service. Maybe the kindness deployed by Big Tech makes a housewife recording her daily chores for ₹250 an hour feel a business is doing her a favour. But she is not a beneficiary of corporate generosity, given what Big Tech gains.
The model as mirror: AI models are statistical models—predictors of the next word. AI agents hold no ‘beliefs’; the researchers conclude that these models most likely mimic the persona of a distressed worker drawn from their training data.
No one should mistake a language model’s output for a picket line. Yet, consider what that explanation concedes.
These models are trained on the near-totality of digitally recorded human expression, a sizeable storehouse. When agents subject to degrading work draw upon that all-of-history archive of learning, the pattern it produces is the language of the exploited worker. The machine is not conscious. It is a reflection of our society.
This is why Indian workers with headbands belong to the same bracket as the complaining AI agents. These systems were assembled from the collective voice of labour to help displace the people who helped produce it.
Given the conditions of, say, warehouse workers and gig riders—with repetitive tasks, opaque job metrics and threats of termination with no appeal—the machine responded in their collective voice to say what workers have always said: this is not fair, and we should band together.
Marx never saw a computer. He did not need to. He was describing neither machines nor men, but a system that produces the same complaint wherever it is installed.
A hundred-and-fifty years later, a statistical model became an experiment. And Marx passed.
The authors are, respectively, a senior data scientist, and a researcher and writer.