Ever since I worked with Mangrove Education, a grassroots teacher-training effort in Indonesia, I've been quietly obsessed with collecting human stories. Its whole premise is disarmingly simple: every teacher is a hero, and every hero has a story. Gathering those stories was slow, noble work. It took us months just to record a handful of interviews — each one a real teacher, a real classroom, a real moment of deciding whether to keep going when the pay was thirty dollars a month and the lights shut off at five.
At the time, that slowness felt like the price of doing honest work. Now I think it was the point.
The AI Industry Ran Out of Internet
Because of where technology is heading, those stories are worth more than we knew. The AI industry has, in a sense, run out of internet. To make agents that act — not chat, but do — companies need something the web can't give them: recordings of real people doing and deciding. Right now, firms are paying workers in India, Nigeria, and beyond to strap cameras to their heads and film themselves cooking, cleaning, working a factory line. “There's no internet for robot data,” as one roboticist put it — so they're building it, one chore at a time.
Out of that scramble came two ideas I can't stop turning over: reinforcement-learning environments — safe places where an AI practices a task and gets scored — and synthetic environments, believable simulations that multiply the practice. The whole game is teaching machines judgment in situations that carry stakes.
We Just Called It Storytelling
And that's when it landed on me, as a producer. The stories we build for television, animation, and film are already this. A good script is a world with stakes: characters who want something, choices that cost, consequences that follow. We have spent entire careers constructing believable human environments. We just called it storytelling.
An environment, to be useful to AI, needs two qualities: it has to be believable, and it has to be measurable. Producers are masters of the first. The second — shaping a story so that a decision inside it can be scored, the way modern AI-agent testing tools do — is a craft we haven't learned yet. But we could. In fact, I've started building a small tool in exactly that direction: one that treats a story like an environment, where the branches, decisions, and outcomes are structured well enough to be tested, not only felt.
A New, Responsible Job for Storytellers
I don't think any of this replaces storytelling. I think it's a new, responsible job for the people who tell stories.
As machines learn to act in the world, someone has to build the worlds they learn in — and keep those worlds honest and human. That has always been our work. It just matters more than it did back when we were coaxing a few teachers in Indonesia to believe their stories were worth recording.