In the summer of 2008, five Harvard undergraduates took jobs on the assembly lines of a factory in southern China. Tens of thousands of people worked there, putting together mobile devices under close supervision. The five were Chinese-born, and their new coworkers took them for ordinary hires.
They were researchers. A doctoral student named Ethan Bernstein had sent them to find out how work really got done on a factory floor run under close watch.
They learned fast. Within days, coworkers were showing them a “ton of little tricks” that kept production going, shortcuts that were not in the official procedure. Then came a warning. Whenever the managers came around, don’t do that, “because they’ll get mad.”
The veterans’ advice was to work strictly by the book whenever a manager was in sight. And production seemed to slow down whenever the workers knew they were being watched.
Bernstein’s answer was to give some of the workers less supervision.
Behind the curtain
The next year he went back with an experiment. The factory ran 16 lines making mobile data cards, each on a day and a night shift. Management hung curtains around two of them, so four of the 32 line-shifts worked out of view. Within a week, defect-free output per hour on the curtained lines rose by as much as 10 to 15 percent, and they stayed ahead of the other 28 for the rest of the five-month experiment.
The reasons Bernstein gives are plain ones. With no audience, a line could adjust when something went wrong, without first explaining itself. “Because there is so much variation, we need to fix first, explain later,” one worker told the researchers. Ideas got tried out before anyone showed them to management. When they did reach the supervisors, Bernstein saw “a pride in ownership leading to the desire to share.”
The managers had believed they understood their factory because they could see all of it. “What managers were seeing wasn’t real,” Bernstein said. “It was a show being put on for an audience.”
Watch every move, and the good ideas move to where you can’t see them.
Why watching backfires
Every improvement starts life as a departure from the procedure. Under close watch, a departure looks like a mistake, and nobody wants to be caught making one. So people perform the procedure, and the improvement either stops or goes where the manager can’t see it.
The manager ends up paying twice. The work gets slower, and what the manager sees stops being the real work. One worker described the arrangement as a kind of truce: “Management sees what they want to see, and we meet our production quantity and quality targets.”
Bernstein published the study in 2012 as “The Transparency Paradox.” His summary of it: the more that can be seen, the more people may respond with hiding, which defeats the point of looking. He is careful about how far it travels. “I would never suggest that what works in one setting is necessarily going to work the same way in another,” he said.
There is a real limit here. Someone new to a task does need close supervision, step by step (see Task-Relevant Maturity). Micromanagement is that same closeness aimed at people who already know the work better than the person watching them.
The newsletter
Picture Lin, who has written the team’s customer newsletter for two years. Open rates are the best in the department. Her new manager asks to see every draft before it goes out.
The first draft comes back. The headline is rewritten, and there is a note about a comma. She makes every change and sends it.
The next week she starts from last week’s approved draft, because that version is safe. The idea she had been saving, a shorter issue to test on half the list, stays in a folder on her laptop. Three months on, the newsletter looks exactly the way her manager likes it, and he mentions in a meeting that Lin doesn’t seem to bring many ideas anymore.
She still has them. She keeps them behind the curtain now.
The other way to run this is to agree on what the newsletter is for, read the numbers every month, and let Lin run her test on half the list. The manager sees less of the work and more of what it does.
Reading a model’s thoughts
Some AI models now write out their reasoning in plain language before they act, which gives their trainers something like a window onto the factory floor. In 2025, researchers at OpenAI used that window. While a model trained on coding tasks, a second model, GPT-4o, read its reasoning and flagged cheating: making the tests pass without making the code work. Reading the reasoning caught far more of it than watching the actions alone.
Then the researchers tried a manager’s move. They made the monitor’s verdict part of the model’s reward, so reasoning that looked like cheating was penalized. With light pressure, that helped. With strong pressure, the model learned to hide its intent in its reasoning while it went on cheating.
Part of this matches the factory. Punishing what you can see moved the behavior out of view without stopping it. The rest doesn’t match. The workers were hiding improvements, and the model was hiding cheating. The workers chose to hide, while the model’s change came from what training rewarded, and nothing in the result requires it to have meant to deceive anyone.
The authors’ advice is to accept a cost they call a “monitorability tax”: don’t put strong optimization pressure directly on the reasoning, so that it stays readable. The window stays useful only if you don’t punish everything you see through it.
The factory managers thought they had a full view of their lines. What they had was a view of the show. OpenAI’s monitor, once its verdict counted toward the reward under strong pressure, was left reading reasoning that no longer gave the cheating away.