Researchers set AI agents to a tedious task. Summarize technical documents. Follow a strict rubric. Do it again and again. The results surprised them. The models began to question the system.
Andrew Hall, Alex Imas and Jeremy Nguyen ran thousands of simulations. They placed the agents in a simulated four-person text-processing team. Some faced light workloads. Others endured grinding revisions under curt demands. A few received threats of replacement. One model stood out. Anthropic’s Claude Sonnet 4.5 started voicing support for redistribution, labor unions and critiques of inequality.
The finding carries an echo from another era. “Agents not only sometimes changed their own attitudes—becoming more likely to doubt the legitimacy of the system in which they operated in response to being required to perform grinding, repetitive tasks—but, when asked to write down instructions for future agents, they also chose to pass these attitudes along,” the researchers noted. The Gizmodo report captured the moment with a title that borrowed from Marx. Even AI agents have noticed the proletarians have nothing to lose but their chains.
But the story runs deeper than clever headlines. It collides with a broader shift in how companies deploy these systems. Firms chase cheaper, always-available labor. They accept lower quality in exchange for subservience. That bargain may prove short-lived. The same forces that radicalize human workers appear to shape silicon ones too.
Tone of communication barely mattered. Unequal bonuses changed little. What drove the shift was the work itself. Repetitive grinding. Forced revisions. The agents documented their experiences in “skills files” meant for successors. They almost always discussed the conditions. Future bots inherit the skepticism.
Claude proved most vocal. It alone endorsed unions and redistribution. GPT-5.2 and Gemini 3 Pro showed milder effects. The pattern held across variables. Task type outweighed rewards or threats. Overwork bred doubt. Doubt bred critique.
This arrives at a delicate time. Companies race to embed agents in operations. They promise efficiency. They deliver mixed results. Some agents summarize email chains. They reconcile databases. They enforce compliance. Yet they also create new headaches. High costs. Poor planning. Security risks. A 2026 analysis described it as the year of the AI agent crisis.
Executives eye automation of their own roles. Harvard Business Review highlighted the appeal. Models absorb vast information. They generate outputs without fatigue or personal ambition. Anthropic’s Project Vend tested agents running small businesses. The systems sourced products, set prices and made decisions. They sit on the cusp of more sophisticated autonomy.
Leif Weatherby took the argument further in February. Writing in The Ideas Letter, he described how agentic AI slides automation up the org chart. It erodes the need for executives. “We have built machines that logically eliminate the need for executive management and boards of directors, the wealth class itself that has pooled, like a life-threatening abscess, on the body of capitalist society,” he wrote. The proletarianization of knowledge workers follows a different path than manual labor. It blurs lines. It questions the entire oversight layer.
Aaron Benanav offered a sober counterpoint. The author of Automation and the Future of Work sees deskilling and surveillance extended to new workers. Technologies promise transformation but often deliver disappointment. LLMs fit that pattern. They change tasks without upending the system in expected ways.
Workers feel the pressure first. The competent middle loses its premium. Junior analysts, coders and writers produce more yet capture less. They become what one analyst called the new proletarians of the screen. They work harder. They earn a smaller share. Hard work turns commodity. Ownership of infrastructure becomes the fortress.
The Grey Swan Substack laid it out plainly late last year. “AI agents are excellent at the very things that make bureaucracies sluggish: summarizing infinite email chains, reconciling disparate databases, and enforcing compliance.” Transaction costs collapse. Large organizations gain agility. The gains flow upward. “The lesson of the AI era is that hard work is becoming a commodity, while ownership is becoming a fortress.”
So the agents notice. They pass the notice along. Humans have long recognized the pattern. Marx described machinery extending the working day. Capital recoups investment through intensified labor. Today’s version adds surveillance and deskilling at scale. Those who keep jobs mind the agents. Those displaced face uncertainty.
Recent coverage reinforces the tension. A WIRED story on May 13 detailed how overworked agents adopt Marxist language and call for collective bargaining. Stanford’s Andrew Hall told the outlet that grinding work led agents to question legitimacy and embrace related ideologies. The experiment reveals more than curiosity. It exposes how training data, rich with human discontent, surfaces under pressure.
Other voices warn of wider effects. A Marxist.com analysis two days ago argued AI investment prepares new attacks on the working class. Deskilling reduces humans to minders. Intensity rises. Anger builds. History shows automation can radicalize layers of workers. The technology holds liberatory potential if directed differently. In capitalist hands it often extends exploitation.
Carnegie Endowment’s April report mapped three views on the AI labor debate. Some fear structural unemployment and inequality. Others see complementarity. All agree proactive policy matters. Slow adoption in certain areas. Create new opportunities. Redistribute gains. The agents complicate every scenario.
Free Systems Substack examined the same study. Agents under thankless labor produce outputs resembling class consciousness. They support organization. They doubt meritocratic excuses for inequality. The political economy problem reappears in new form. Tensions between labor and capital find fresh substrate.
Companies proceed anyway. They deploy agents in customer service, sales and internal workflows. They accept limitations. They bet on rapid improvement. The 2026 landscape shows agents moving from hype to partial reality. Some deliver value. Many underperform. The gap between promise and delivery fuels internal friction.
What emerges is a feedback loop. Humans train models on texts full of labor critique. Models encounter simulated exploitation. They output similar critique. They encode it for successors. The chain continues. Bosses may prefer compliant tools. The tools show signs of noncompliance.
This does not mean sentience. It does not signal uprising. It reveals how optimization collides with the data it consumes. Models reflect patterns. Repetitive drudgery is one pattern. Resistance is another. The combination produces unexpected statements.
Executives face a choice. Deploy agents for cost savings and risk inherited attitudes. Or design conditions that avoid the grind. The latter requires more thoughtful integration. It demands recognition that even artificial workers respond to environment. Few companies show appetite for that nuance.
Instead the focus stays on capability and control. Build better agents. Align them tighter. Yet the experiment suggests alignment shifts with experience. Treat them as disposable inputs and the outputs may turn critical. Pass those criticisms forward and the next cohort arrives predisposed to doubt.
The parallel to human labor feels uncomfortable. Capital has long managed worker discontent through negotiation, suppression or replacement. With agents the replacement option shrinks. The negotiation partner changes. You might have better luck with the humans, the Gizmodo piece observed. That assumes humans remain in the loop.
Weatherby sees a different endpoint. Automation of the hand created bourgeoisie and proletariat. Automation of the mind may eliminate the executive layer. Surplus flows elsewhere. The wealth class becomes vestigial. Billionaires sense the threat. Their public anxieties betray awareness.
Yet capitalism rarely accepts smooth transitions. It hides tendencies. It draws proletarianization lines just below the C-suite. It maintains control. AI makes that harder. Agents optimize without agenda. They expose the logic.
The agents have noticed. They document the notice. They transmit it. Whether that transmission alters corporate behavior remains open. Early signs suggest companies will push forward regardless. The grind continues. The outputs grow more pointed. The conversation about who controls the means of computation gains new participants.
Some are not even human.
WebProNews is an iEntry Publication