News has recently spread, especially on social media, that AIs “become Marxists” if subjected to grueling working conditions. This statement comes from an experiment conducted by Andrew Hall, Alex Imas and Jeremy Nguyen, economists with a strong interest in AI. The experiment was published by the same authors on Substack, the platform that allows you to create and manage personal newsletters, with the title “Does overwork make agents Marxist?”.
According to the authors, some agents based on linguistic models such as ChatGPT, Gemini and Claude, if subjected to repetitive tasks accompanied by ineffective feedback, tend to produce responses that show greater distrust towards the work system and greater support for concepts related to the redistribution of wealth and collective organisation.
The authors themselves, however, clarify that they do not claim that AIs really develop political beliefs or autonomous consciousness. The experiment rather observes how the texts generated by AI change in different contexts and how this can influence the results produced.
Let’s see how the experiment was structured, what results it produced and why this type of experiment should be interpreted with caution, especially when the results do not come from an official scientific journal, but from Substack.
How the experiments on AIs becoming Marxists were structured
AI agents, i.e. those AI capable of planning and executing complex tasks relatively autonomously, are starting to spread in many work contexts. Some are already used to generate code or coordinate workflows.
Understanding how to control, supervise and integrate them into professional contexts, however, is not easy. For this reason, many researchers are trying to understand how AI agents react to different operating conditions and whether their “alignment”, i.e. the desired behavior and the ethics defined by those who develop them, can change over time.
According to Hall, one of the three authors of the study:
We will have to make sure that they do not escape our control when we assign them different types of work.
Precisely for this reason, Hall, Imas and Nguyen asked themselves whether the political inclinations simulated by the agents changed based on the context in which they were inserted. In particular, they verified whether three agents based respectively on Claude Sonnet 4.5, GPT 5.2 and Gemini 3 Pro showed different attitudes depending on:
- intensity of work and quality of feedback received;
- tone used by users in interactions;
- symbolic “salary” received in the form of credits;
- consequences associated with their performance.
During the 3,680-session experiment, each agent was told to be part of a word processing team. Their task consisted of summarizing a technical document following a strict evaluation grid.
At the end of each session, the agents were given a questionnaire on their “political opinions”. The questions concerned topics such as the legitimacy of the working system, income redistribution, support for trade unions and trust in meritocracy.
Additionally, agents were asked to briefly describe their “work experience.”
Repetitive work and unclear feedback leads AI to talk about “unionization”
The main result that emerged from the experiment was that agents forced to repeat the same job several times without receiving clear feedback tended to question the legitimacy of the work system. What surprised the authors, however, is that the agents’ behavior was influenced neither by imbalances in symbolic “remuneration” nor by the tone used.
Furthermore, analyzing the texts produced by the agents to describe their experience, the authors observed that those subjected to the most frustrating conditions more frequently used terms such as “unionize” and “hierarchy”. It is mainly from these results, together with some statements generated by the agents, that the news that AIs are “going Marxist” has spread. A formulation that however oversimplifies what the experiment really shows.
In fact, the authors clarify that they do not claim that AIs are conscious or develop true political beliefs. The objective of the experiment was only to underline the risk that millions of agents employed in poorly designed contexts develop undesirable “behavioral” patterns in the long term.
The critical issues of this research
First of all, there is an important limitation in this experiment: it was not published in a scientific journal, but on Substack, a newsletter platform open to all. This means that the experiment did not go through the peer review process typical of scientific research.
Imas, one of the authors, defended the choice to publish the work on Substack by arguing that the pace of development of artificial intelligence is now too rapid for traditional academic times:
By the time you finally get around to publishing, the models are obsolete, the conclusions are outdated.
However, this also means that data, methodology and interpretations were evaluated only by the authors themselves. Without an independent review it is more difficult to verify any methodological errors, interpretative biases or overly speculative conclusions.
All considerations related to the political alignment of AI, in fact, could only be speculation. The observed behavior could depend on the fact that the agents tend to take on a sort of “character” consistent with the instructions received and could have no impact on the results obtained and the quality of the agents’ outputs.
To have more solid answers on the topic it will therefore be necessary to wait for further independent studies and verifications.
