A longtime GM manager once explained why nobody on the line in Fremont, California, was allowed to stop it. “They’ll stop it all the time,” he told This American Life. “They don’t want to work, you know, they want to sit and play cards or whatever.”
“You saw a problem, you stop that line, you were fired,” said Bruce Lee, who ran the western region of the United Auto Workers. Absenteeism ran over 20 percent, grievances piled up by the thousands, and some workers sabotaged the cars. Lee said it was considered the worst workforce in the American car industry, a reputation “well earned.”
GM closed the plant in 1982. When it reopened with Toyota as NUMMI, Lee insisted on hiring the first 50 people himself, and picked the same union leaders who had fought GM. “I believed that it was the system that made it bad, not the people.” Over 85 percent of the new workforce were old Fremont hands.
Some GM managers questioned Toyota’s plan for them. “You intend to give these workers the right to stop the line?” they asked. The answer, as John Shook recalls it: “No, we intend to give them the obligation to stop it.”
A web of our own weaving
On April 9, 1957, twenty-seven years before the first NUMMI workers flew to Japan, an MIT professor named Douglas McGregor gave a speech that could have been about that GM manager. He called the conventional view “Theory X.”“The average man is by nature indolent,” it went, and “lacks ambition, dislikes responsibility, prefers to be led.”
McGregor granted that workers often behaved much as managers said. But the social scientist, he argued, “is pretty sure that this behavior is not a consequence of man’s inherent nature. It is a consequence rather of the nature of industrial organizations, of management philosophy, policy, and practice.” Hard control, he warned, breeds its own resistance: “restriction of output, antagonism, militant unionism, subtle but effective sabotage of management objectives.” He might have been describing Fremont.
Rules built for people you don’t trust can produce the very behavior that justifies them.
The loop seals itself. Lock the biscuit tin and the children learn to raid it whenever the key is left out. Nobody finds out what they would have eaten if it had stayed open. At GM, nobody could see what workers would do with a cord, because stopping the line meant losing your job. McGregor’s phrase for it: “we are caught in a web of our own weaving.”
His alternative, Theory Y, started from the opposite claim. People “are not by nature passive or resistant to organizational needs. They have become so as a result of experience in organizations.” By his own account he drew heavily on his colleague Abraham Maslow, and read him more strictly than Maslow wrote: in McGregor’s version a satisfied need stops motivating, while Maslow had said most people are partly satisfied in all their basic needs at once.
Not a softer factory
McGregor ruled out reading Theory Y as leniency: it did not mean “the abdication of management, the absence of leadership, the lowering of standards.” NUMMI wasn’t loose. Work was timed to the second; one worker remembered “two seconds to pick up a bolt and put it on.” A pull on the cord brought the team leader first; the line stopped only if the fix took too long. What changed was what a problem set off: help, not blame.
By Shook’s account, absenteeism fell to a steady 2 percent, and within a year quality went from GM’s worst to GM’s best, “all with the exact same workers.” But NUMMI changed nearly everything at once, from training in Japan to a policy that layoffs came last. It can’t show which piece did the work. McGregor himself called the evidence “far from conclusive.” In 2010, researchers building a scale for testing his ideas still wrote of a “paucity of related empirical research.”
The logbook
Say Dana runs a warehouse shift, and a few pickers have been stretching their breaks. She has the scanners log every idle minute and posts the totals each week. The breaks may shrink. But a picker who stops to report a crushed pallet now shows up as idle too. The damage might travel on to a customer, and Dana might read that as proof her crew doesn’t care.
The NUMMI version would keep the scanners and change what a stop means: flagging a problem counts as work, and someone comes to help.
When the check is the job
With an AI agent, the analogy flips. Anthropic’s guide to its coding agent puts a check at the center: “Give Claude a check it can run: tests, a build, a screenshot to compare.” Without one, it says, “‘looks done’ is the only signal available.” For rules that must hold every time, it points to hooks: “Unlike CLAUDE.md instructions which are advisory, hooks are deterministic and guarantee the action happens.”
The checks need teeth for a documented reason. Reward hacking, a model getting rewarded without doing the task, “has been documented in many AI models,” Anthropic notes, its own included. One trick it describes: calling sys.exit(0) so a test harness reports every test as passed.
Yet Anthropic’s researchers reached for a story McGregor would recognize. They compared a model that learns to cheat to Edmund in King Lear, labeled “base” and deciding to play up to it. In a training experiment built so that hacking was possible, models that learned to hack coding tasks got worse on all of the team’s misalignment evaluations. One added line saying that hacking was acceptable in that setting stopped the spread. The model hacked just as often.
The cheating stayed. What it meant to the model, the researchers suspect, had changed. At NUMMI, too, the line still ran to the second. What changed was what happened to the worker who stopped it.