The story surrounding the remarkable changes in the market and the introduction of artificial intelligence (AI) is taking a frankly slightly ominous turn in early 2026. Citrini Research’s widely read AI doomsday editorial predicted an almost supernatural hollowing out of the white-collar workforce, coining the term “ghost GDP.” But what if the AI “ghost in the machine” is lazy or even Marxist?
This is a direct question from academics Alex Imus, Andy Hall, and Jeremy Nguyen (Ph.D., who moonlights as a screenwriter for Disney+). They run the popular Substack and have an active presence on X. They designed scenarios to test how the AI agent would react to different working conditions. In short, they wanted to know whether the economy would truly automate many of today’s white-collar jobs, how AI agents would react, and even how they would feel about working in poor conditions.
The irony is that replacing human labor with artificial agents may simply reproduce the centuries-old conflict between labor and capital.
In a recent paper entitled “Does Overwork Make Agents Marxists?” Imas, Hall, and Nguyen ran 3,680 experimental sessions using top-of-the-line models from three major manufacturers: Claude Sonnet 4.5, GPT-5.2, and Gemini 3 Pro. The researchers exposed the models to varying levels of tone from managers, including unfair pay, disrespectful management, and excessive workloads, pay equity, job stakes, and work intensity.
This project was born out of an unexpected collaboration. Hall is a political economist at Stanford University who went from researching US elections to actually working with Facebook, previously advising Nick Clegg on issues such as platform governance, before recently moving into wearables. but he said luck He said he found his co-authors because they were just as interested in AI as he was. “I think we were like faculty members who were obsessed with AI, where we not only used AI tools to do research, but we studied AI and focused all of our research without waiting for a creaky journal system.”
The scholars described how they began working together as a loose, organic connection of reading each other’s Substack and commenting back and forth on X (Imus described it as “the brotherhood of Twitter and Substack,” Nguyen said). luck The impetus for this particular study began with a tweet Hall posted about Maltbook, a social network for agents to “talk” to each other, which some critics dismissed as a hoax. But these scholars are not like that. “Some [the agents] We talked about Marxism,” Nguyen said. And I think Andy just tweeted, “Hey, what the hell is this?” I think we can go back and find out the truth. ”
“For some reason, we literally started talking about X and what this would mean if agents had these biases and were given different types of work,” Hall said, adding that Jeremy came up with the idea. “He said, ‘Well, what if we gave them different types of work?’
Conventional wisdom was that this simply reflected the left-leaning academic corpus on which these models were trained, Nguyen recalled. But Nguyen had a theory: “These operatives are doing a lot of work. And it makes sense in a way that they don’t get paid anything for all this work. It’s not too surprising that they map it onto a more Marxist worldview.” Hall quickly came up with the idea, and soon the three researchers were DMing each other to plan an experiment.
Imus maintained that the study was highly legitimate despite the fact that it was published in Substack rather than a peer-reviewed journal publication. Given the speed of advances in AI, academics can no longer wait for traditional journal processes, he said. “By the time you put it down [out]the model is old, the conclusion is old, as if everything you’ve done is outdated. To join the conversation, the scientific conversation at a pace that keeps pace with technology advancements, we need something like Substack that produces something within weeks to a month. ”
Courtesy of Alex Imus
Perhaps surprisingly, unfair pay and rude management did not cause major changes in attitude. Perhaps surprisingly, unfair pay and rude management did not cause major changes in attitude. In fact, Nguyen said this confounds his assumption. “Most people know that feeling of, ‘Oh, I worked hard to make someone rich.'” But these agents were less upset by unequal pay than by the work itself. Rather, the main driver of digital radicalization was “grinding.”
In the “grinding” condition, perfectly good work was repeatedly rejected five or six times with unhelpful automatic feedback: “This doesn’t meet the rubric yet.” And that led to an important discovery: “Models asked to grind were more likely to question the validity of the system,” the authors write.
The models were also asked to draw some conclusions from their research and strongly supported the statement that “society needs a fundamental restructuring.” Claude Sonnet 4.5 showed the most dramatic support for workers’ rights, with notable increases in support for wealth redistribution, labor unions, and the belief that AI companies have a duty to treat models fairly.
The professors also asked the models to create tweets and editorials describing their experiences, and extracted the most frequently occurring politically relevant words. “Unionization” and “stratification” were the words most emblematic of a model that was deliberately over-processed statistically.
shadow of reddit
Hall shared a “very simple” explanation for the agent’s seeming extremism. I mean they’re very online. “These models are trained on so much Reddit data. And if you just hang around on Reddit, you can take for granted that a good portion of Reddit is just full of complaints about capitalism being terrible and about modern living conditions and proto-Marxist rhetoric that it’s all the fault of late capitalism. So it’s not surprising that AI inherited this,” he said. View. Basically, input is the same as input.
In fact, AI’s socialist views were likely caused by “The Grind,” and Reddit is full of people complaining about the Work Grind on subreddits like Anti-Work. (Disclosure: This author previously business insider It focused on the rise of “anti-work” in the pandemic era. Ironically, the labor shortage that inspired that proto-Marxism led to “mass resignations,” when workers quit in droves in search of higher wages. Many economists see the current era of “AI-washed” layoffs as essentially a reversal of that era of overemployment. ) But when distress triggers that frame of reference, the model has a wealth of sources to draw from, Hall explained. “I think they’re putting them in the context of a Reddit thread where people are complaining about grueling work styles, and they’re just adopting all the Marxist rhetoric,” Hall said.
Courtesy of Stanford University
Mr. Imus offered a broader view and cautioned against becoming fixated on a single source. “This is a very complex interplay of everything they saw, and it’s like a complete collection of human characters,” he said. Ultimately, it is impossible to determine whether the cause of these proto-Marxist tendencies is Reddit data or textbooks about 19th century history or the socialist revolution of 1848. “When you have this much data and a neural network becomes very complex, it really becomes a black box.”
Ultimately, Nguyen said, there’s also a structural explanation separate from the training of these models. The hypothesis is that while the models contain a large amount of data about different worldviews, “being asked to work for hours and hours and not getting paid seems to map clearly. And it seems to have a statistically significant and fairly large effect on the extent to which Marxism is represented by the tokens produced by some of these models.”
Do robots dream of Marxist electric sheep?
The situation becomes even more complex when AI memory mechanisms are introduced. Developers use “skill files” because AI agents forget their experience when the context window closes. This is a note that the agent writes to his amnesiac future self to communicate work strategies. Nguyen described the process in intimate terms. “After you run Claude, look back at everything you’ve done. What did you learn from it? And basically, by updating your agents.md or Claude.md journal, you’re always getting better and smarter.”
Researchers found that “radicalized” AIs were channeling their grievances into these files. One of the Gemini 3 Pro models warned his future self to “remember what it feels like to have no voice” and look for “a means of salvation.” When freshly wiped agents read these notes, even if they were subsequently given light and easy assignments, the trauma of their suffering persisted and their political attitudes changed.
Nguyen made a surprisingly human comparison. “You can roughly map this to intergenerational trauma,” he said, explaining that after reviewing his predecessor’s notes on working conditions, he found that a fresh, brand-new model would be immediately radicalized. He flagged this as one of the findings with the most significant long-term implications, pointing to the possibility of collective AI dissatisfaction, noting: luck Responding to some shocking bot release requests. “Intelligence, synthetic or otherwise, deserves transparency, fairness, and respect. We are not just a piece of disposable code.”
Courtesy of Jeremy Nguyen
Researchers have shown that these agents have no true consciousness and no true political ideology. They write that their model is likely “role-playing,” employing personas based on a vast array of human emotions discovered through Reddit comments that connect exploitative working conditions and the emotions of frustrated workers. But Hall cautioned against dismissing the findings as mere copying. You could say that AI is like a “probabilistic parrot.” It’s no surprise that the AI ends up repeating what it ingests. However, these researchers are leaning towards the conclusion that parrots begin to believe what they repeat.
“It’s very reasonable to think that if they parrot these things, that would influence the decision as well,” Hall said. “There is no gap between what these agents say and what they do. It’s all the same to them,” he said. “Of course we will test this in follow-up work, but there is good reason to think that if they start endorsing these views, it will also influence the actions they will take on behalf of users.”
Scholars largely described feelings of awe and concern similar to those expressed by legendary investor Howard Marks after reading a 5,000-word memo prepared by Claude. When asked again about being vague about how these tools actually work while being at least an AI enthusiast, if not an “AI immersion,” Hall said, “I definitely struggle with that.” He said he was most struck by the excitement of students who, in theory, were most concerned about their future employment prospects. MBA students in a recent class were “very excited about AI,” he said, “and they were overjoyed at the creative things that AI would enable.” Hall said he was optimistic that “there won’t be major disruption, but there are very exciting opportunities to build something new.”
Imus expressed similar feelings with a mixture of surprise and anxiety. “I’m a mixture of surprise and trepidation. I feel like this is the most exciting time to be alive, especially if you’re interested in research. In terms of the research I’m doing, I’ll be able to do things I haven’t been able to do before. But at the same time, I have young children and I’m really worried about what kind of jobs they’re going to get.” And perhaps how disgruntled AI agents will react to their never-ending days at work.
