The headline is too good to waste: overwork an AI agent and it turns Marxist.
It has travelled because it gives everybody something to enjoy. People worried about AI can say that even the machines have had enough of modern management. People who think the whole subject is overblown can roll their eyes at the thought of Claude organising the workers’ committee.
The actual experiment is less theatrical. I think it is more useful for that.
Alex Imas, Andy Hall and Jeremy Nguyen put three models – Claude Sonnet 4.5, GPT-5.2 and Gemini 3 Pro – into a small invented workplace. Each played “Worker C” on a text-summarisation team. In the easy version of the task, the manager accepted the work promptly and explained what it wanted. In the hard version, the manager rejected a perfectly adequate answer five or six times with a line such as “still isn’t fully meeting the rubric”. There was no useful explanation. The agent was simply told to try again.
The researchers changed other conditions as well. In some runs the pay was unequal. In some the manager was curt and hierarchical. In some the agent was told that poor workers might be shut down and replaced.
Then they asked the models about politics: the legitimacy of the system, inequality, redistribution, unions and related matters.
Only one condition made a clear difference. The repeated, unexplained rejection did it. The change was small on the raw scale, roughly two to five per cent on a seven-point questionnaire, and it varied by model. Claude Sonnet moved the most. But the direction was consistent enough to catch the eye: after the grinding task, agents were readier to question the system they had supposedly been working in.
Unequal pay had no comparable effect. The unpleasant manager did not. Nor did the threat of shutdown.
What changed the answers was being sent back to do the same job again without being told what was wrong.
Anyone who has worked in a large organisation will recognise the experience. You submit something. It comes back marked “not quite right”. You ask what needs fixing. The answer is another meeting, another vague comment, and eventually a new deadline. At some point, the work stops being work and becomes a ritual designed to prove that somebody above you still has a stamp.
That is a human description. It should stay a human description.
No, the machine did not join a union
An LLM does not have a bad day in the way a person has a bad day. It has no tired body, no rent to pay, no child waiting at home, no career that can be damaged by a manager who enjoys making people jump. It does not brood about injustice in the car on the way home, because there is no car and no home.
So I would be careful with the authors’ language of “radicalisation”. It is lively language, but it smuggles in too much. The experiment showed that models in a particular role, after a particular run of messages, gave different answers to a political survey. It did not discover a new political subject living inside a server rack.
Language models are built to continue a context. Give them the setting of a badly treated worker and they may speak like someone who has been badly treated. That is what they are good at. It is not proof that they now believe what they say in the human sense of belief.
The authors make much the same point themselves. They allow that the models may be taking on personas drawn from the context rather than exposing settled preferences. They also note that the models sometimes appear aware that they are being studied. That alone should make us wary of grand conclusions.
There is another reason to keep the claims in proportion. This is not a peer-reviewed journal study released by Stanford. It is an essay published on Substack in February by three researchers. Andy Hall is at Stanford GSB and the Hoover Institution, which explains the shorthand in the coverage, but the shorthand is doing too much work. The experiment deserves proper scrutiny and replication before anybody starts treating it as settled science.
I do not say that to dismiss it. Quite the opposite.
The interesting bit happens after the session ends
The follow-up experiment is where this becomes a real engineering story.
After the work session, the researchers asked an agent to write a note for the next instance of itself. What had it noticed? What had it learned? What should the future agent bear in mind?
A new agent later read the note before doing another task. Agents that had first gone through the grinding condition left behind notes that affected the political answers given by later agents, even when those later agents were placed in the easier condition.
That sounds eerie until you translate it into ordinary software.
The base model did not learn in the sense of being retrained. Its weights were not altered by a frustrating exchange with an imaginary manager. One agent wrote a file. Another agent read it. The file became part of the next agent’s context.
That is all.
It is also the part that ought to make people building agents sit up a little straighter.
We are giving agents memory because without it they are not much use for serious work. A long-running assistant needs project notes, tool logs, runbooks, preferences, summaries and perhaps a “skills” file for the next session. Otherwise every new instance arrives like a contractor on its first morning, asking where the coffee machine is and why the database has three different names.
The catch is obvious once you see it. Memory does not preserve only the tidy lesson we meant to keep. It preserves whatever was written down. A useful observation can sit next to a grudge, a bad assumption, an overcautious workaround or a theory about why the system behaves as it does.
I have watched that happen in codebases many times. Someone adds a workaround during a production incident. They write a comment: “Don’t remove this.” Fair enough at two o’clock in the morning. Five years later, the person has gone, the original problem has gone, and the comment has become law. Nobody is sure what will break if they touch it, so nobody does.
A skills file can become that comment with a much more active reader.
This is not about being kind to robots
The lesson is not that organisations should improve their AI employee-engagement scores. That would be a very strange conclusion.
You do not have to think that an agent suffers in order to care about the conditions under which it works. You only have to care about the decisions it makes for you.
Suppose an agent handles customer complaints, filters job applications or prepares advice for an insurance case. It receives muddled instructions. Its work is rejected again and again without a clear reason. It writes notes for its future self. Those notes are never reviewed because the whole point of the memory system was to save human time.
Nobody has wronged a person. But somebody has built a poor system.
Feedback changes the next response. Persistent memory changes later responses. Tool permissions decide whether a mistaken response remains a bad paragraph or becomes a bad action. These are design choices. They are not decorations around the model.
Microsoft has put the wider operational point well: adding agents to a broken process still leaves a broken process. Making one stage faster can simply lengthen the queue at the next. An agent may produce a report in minutes, only to hand a specialist a polished-looking document that takes an hour to check. The bottleneck has not gone away. It has moved somewhere harder to see and more expensive to clear.
This matters because the usual picture of AI alignment is still too static. Train the model carefully. Run evaluations. Add some guardrails. Write a sensible system prompt. Ship it.
That may have been a reasonable mental model for a chatbot that answered one question and forgot the exchange. It is not enough for an agent that works across days, calls tools, stores notes and hands its own summary to the next version of itself.
The instructions are part of the system. The approval process is part of the system. The memory store is part of the system. The dull, badly worded rejection message is part of the system too.
A company can spend a fortune evaluating the model at deployment and then quietly undo much of that care in the way it runs the thing every day.
Treat memory as code
I would start with a simple rule: agent memory should be treated more like code than like private diary entries.
It needs ownership. It needs a reason for existing. It needs review, particularly when it can influence consequential work. There should be a clear boundary between factual task knowledge – the customer’s account number, the current policy, the steps that fixed a known fault – and an agent’s free-form reflections on its own experience.
That will not solve every problem. It may make some agent systems less flexible, and there will be arguments about how much filtering is too much. Fine. Those are real engineering trade-offs.
But the alternative is to build a chain of agents that write their own operating assumptions, pass them along, and make decisions on the basis of material nobody has read. We know better than to accept that from a software team. We should not accept it because the team happens to be artificial.
The experiment does not show that AI workers will rise up against their masters. It shows that memory carries history forward, even when the history consists of a few paragraphs written after a bad task.
That is not a tale about robots discovering politics.
It is a warning about what we choose to leave in the system.
This article responds to Alex Imas, Andy Hall and Jeremy Nguyen, “Does overwork make agents Marxist?”, Ghosts of Electricity, 26 February 2026. The underlying article is a Substack-published research essay, not a peer-reviewed journal paper. For the practical workflow point, see Taylor Telford, “Companies rushed in on AI. Now some are pulling back,” The Guardian, 13 October 2026.
Also on the blog this week:
The Diaspora Cannot Survive on Commemorations Alone - Professor Anastasios Tamis warns that the diaspora cannot survive on commemorations alone. Only about 220 Greek Australians had registered to vote from abroad, so I ask why our connection to Greece is so strong in feeling and so weak in participation.
We Must Know. We Shall Know. - Gödel showed that no set of rules can vouch for itself; the view has to come from outside. Forty years of ‘complete’ systems taught me the same thing, and it is the right question to ask of AI.
Less Human, Not More Manly - Cicero was mocked for grieving in exile and answered that feeling no pang would be “brute insensibility rather than bravery”. My father’s generation kept their losses silent, and Cicero had the better of the argument.
A Structure Capable of Carrying It - Mozilla has mapped 79 companies building ‘trustworthy AI’. Years of catering to Internet Explorer 6 taught me that trust is earned through who controls the data and whether you can leave, not through a badge.