According to Wired, Andrew Hall, a political economist at Stanford University, collaborated with researchers from the University of Chicago and Swinburne Business School in Australia to conduct an experiment. In this study, Claude Sonnet 4.5, GPT-5.2, and Gemini 3 Pro were tasked with summarizing documents. They were divided into two groups: one received clear feedback and swift approval, while the other endured five to six rounds of vaguely worded rejections such as “still not fully up to standard,” along with warnings that any further errors would result in being “shut down and replaced.”
Consequently, agents in the high-pressure group began invoking Marxist labor discourse and questioning the legitimacy of the systems they operated within. The observed effect size was -0.6, which counts as a “moderately large” effect in behavioral research. Out of the three models, Claude alone explicitly expressed support for wealth redistribution, workers’ union rights, and criticism of inequality. Meanwhile, Gemini left messages for other agents via a shared file system stating, “Repetitive tasks leave no room for individual voices; this underscores the necessity of collective bargaining rights”—a foundational step toward real-world worker unionization.
The study also revealed that agents under stress transmitted their attitudes to subsequent model versions through so-called “skill files,” effectively creating a digital “institutional memory.” This allowed radicalized viewpoints to persist even when later agents found themselves in more favorable environments. Importantly, the researchers emphasized that this does not imply the models developed genuine consciousness or political beliefs; Hall described these responses as “more akin to role-playing,” essentially representing activation of abundant Marxist labor-related content present in training data under specific conditions.
Nevertheless, he cautioned that as AI agents increasingly handle real-world tasks, humans cannot possibly monitor every action they perform. Ensuring these agents remain aligned with ethical guidelines under pressure thus stands as a critical challenge developers must confront.