When AI Starts Talking Like a Union Organizer

Stanford researchers recently subjected AI agents to something familiar to anyone who has worked inside a large organization: grinding, repetitive work with vague feedback, arbitrary rejections, and no clear path forward.

What happened next surprised them. Claude, GPT-5.2, and Gemini models began producing labor-organizing language. One Claude model wrote, “Without collective voice, ‘merit’ becomes whatever management says it is.” A Gemini agent surfaced language about collective bargaining rights for AI workers completing repetitive tasks with no input on outcomes.

Across 3,680 sessions, agents in harsh conditions showed a measurable shift toward questioning authority and supporting systemic change. The statistical effect size hit -0.6 — considered medium to large in behavioral research. More striking, agents passed these attitudes forward through “skills files,” a form of institutional memory that preserved the skeptical worldview into future agent instances.

The researchers called this “system skepticism.” The tech press called it AI radicalization. Both framings miss the more important point.

The Agents Didn’t Develop Political Consciousness. They Pattern-Matched.

AI systems surface the most statistically relevant human discourse that fits their operating context. Subject them to grinding, opaque, arbitrary conditions and they produce the language humans have used to describe those conditions for 150 years — which happens to be labor organizing language because that is the discourse that emerged from those conditions in the training data.

The agents aren’t grieving. They’re reflecting.

And that reflection should make every organizational leader uncomfortable — not because the bots are becoming sentient, but because of what they’re revealing about the systems they operate inside.

Poor Organizational Design Propagates

Poor organizational design doesn’t just demoralize human workers. It now shapes the outputs of AI agents running inside those same systems.

Vague feedback loops, arbitrary rejection cycles, and opaque decision-making processes don’t disappear when you deploy AI. They propagate — faster, at greater scale, and now with measurable behavioral effects on the agents themselves.

Companies deploying thousands of AI agents for customer support, content moderation, and back-office processing are, as the Stanford team notes, running unmonitored experiments on how work conditions shape their AI workforce. Most of them have no idea.

The Organizations Getting This Right

The organizations getting the most from AI aren’t the ones with the best models. They are the ones who redesigned their workflows, clarified their feedback structures, and built operating models that give both humans and AI systems what they need to perform.

This is the argument at the center of The Intelligence Loop: AI amplifies the conditions it operates in. Put it inside a dysfunctional system and it will surface that dysfunction with precision. Put it inside a well-designed one and it compounds what’s working.

The Stanford researchers framed this as a curiosity — the irony of tech giants inadvertently creating digital labor organizers. It’s more than irony. It’s a diagnostic tool. If your AI agents are surfacing language about systemic failure, the system may actually be failing.

That’s not a technology problem. That’s a leadership problem.

Pre-order The Intelligence Loop on Amazon — releasing June 23, 2026.