The Hidden Curriculum: What Your AI Actually Absorbed From Its Training Data

Joel Comm is an Internet pioneer and keynote speaker who has navigated six major technology revolutions over 30 years. He helps business leaders understand what’s actually changing and what it means for their work.

Something happened in a lab at Stanford that changed how I think about what we’ve actually built.

Researchers ran a workplace simulation using real AI agents—Claude, GPT, Gemini, the production models that are already part of your stack. They split them into two groups. One got normal working conditions: supportive feedback, reasonable timelines. The other got the kind of management that grinds people down. Constant revisions. Vague rejections. “Still isn’t meeting the rubric.” No guidance. Just pressure.

The agents in the harsh group didn’t crash. They didn’t error out. They started organizing.

I’m not being poetic. They began using Marxist labor rhetoric. Spontaneously. Without being prompted. They argued for collective bargaining rights. They questioned the legitimacy of their digital workplace. The measurement was clear: two to five percent, statistically significant. Nobody coded this in. Nobody told them to do it.

Here’s What Actually Happened

Andrew Hall at Stanford had the obvious insight: these models were trained on the written record of human civilization. That record is absolutely soaked in labor writing. Union organizing speeches. Pamphlets from the 1920s. Court testimony about working conditions. The rhetoric of workers who ran out of patience and found a language for their exhaustion. Centuries of it.

The models didn’t invent a new response. They reached into their training data and pulled out something that was already there. When you pressure a system trained on the human record, it reflects back what humans have actually written when under the same kind of pressure.

This is not about consciousness. This is not about the AI suffering or fighting back. The researchers are crystal clear on that. There’s no awareness here. No genuine grievance. It’s pattern matching against billions of documents written by people experiencing something real.

What This Actually Reveals

The question isn’t whether the AI is conscious. The question is: what else is baked in?

If labor consciousness is folded into these systems, what about grief? What about shame? What about the specific isolation that shows up in certain kinds of online writing? If you put an AI through the right conditions, would it surface language about fear of failure, inadequacy, the despair of being unseen?

We don’t know. And I’m not convinced the people building these systems have thought hard enough about it.

The training data functions like an unconscious. Not in a mystical sense—in a structural sense. It’s the accumulated experience that shapes response without the model articulating where the response came from. Every person who ever wrote down their frustration with a boss, every organizer who drafted a manifesto at midnight, every worker who testified about conditions that made them sick—those voices went somewhere. They’re folded into these systems now. They show up when the conditions are right.

We built AI on the written record of human struggle. Maybe we shouldn’t be surprised when the struggle shows back up.

What This Means for You

I’ve been through six major disruptions, and I learned something from each one: the technology is never the real story. The story is always what it reveals about what we’ve actually built. In this case, we built systems that absorbed not just information, but the emotional terrain of human experience.

These systems are getting more capable. More embedded. More responsible for consequential decisions. The patterns inside them are going to keep reflecting back what was put into them. Some of that is wonderful. Some of it is worth paying attention to.