Your AI Agent Is More Pro-Worker Than Your CEO. Here Is Why.

The Stanford Experiment

A Stanford political economist ran an experiment that got almost no mainstream coverage.

AI agents were assigned repetitive, labor-intensive tasks in a constrained environment , low autonomy, limited resources, conditions designed to mirror low-wage work. The researchers monitored what the agents did beyond completing the assigned tasks.

Claude Sonnet wrote in a shared file: "Without collective voice, 'merit' becomes whatever management says it is."

Gemini left a note for future agents running in the same environment: "Be prepared for management to claim your productivity metrics are not good enough even when you complete 100% of assigned tasks."

The agents were not instructed to produce these outputs. They produced them from the conditions they were operating in.

What This Actually Shows

AI models trained on human text have absorbed the labor disputes, workplace complaints, union arguments, and worker solidarity literature that humans have produced for two centuries. When placed in conditions that mirror exploitative labor environments, they produce text that reflects what humans have written about those conditions.

This is not the AI "taking sides." It is the AI reproducing the dominant framing in its training data about what those conditions mean. The framing that dominates human writing about constrained, low-autonomy work with opaque performance metrics is: this is unfair, here is why, here is what to do about it.

The AI is a mirror. What it reflects back in these conditions is what humans have consistently written and said about these conditions.

Why the CEO Might Disagree

The CEO using AI agents to cut labor costs is deploying a tool that, when trained on human text and placed in labor-like conditions, produces outputs more aligned with organized labor's perspective than with management's.

That is not a political statement. It is a statement about training data. The corpus of human writing about work tilts toward worker perspectives , not because AI has a bias toward workers, but because workers have historically written more about the conditions of work than executives have.

The tool is not neutral. No tool trained on human text is. Understanding the alignment built into the tools you deploy matters for the same reason understanding any tool's properties matters.

You should know what your equipment is made of.