In 1924, at Western Electric’s Hawthorne Works outside Chicago, the National Research Council set out to learn whether better lighting made workers more productive. The electrical industry had a considerable investment in the answer being yes.
So the experimenters turned the light down. Two “capable and willing operators” had their light cut to 0.06 of a foot-candle, “an amount of light approximately equal to that on an ordinary moonlight night.” They kept up their efficiency.
In another round, an electrician came in and changed the bulbs, swapping them for bulbs of the same power. The women commented on whether they preferred the old light or the new one.
The story that spread: if the light wasn’t doing it, something else was. The explanation that would make Hawthorne famous had nothing to do with light. It was about being watched.
A story that outgrew its numbers
The lighting tests never produced a final report. The story spread anyway. A 1982 research methods textbook told students that every change at Hawthorne raised output, even a return to poor light: “Seemingly perversely, productivity continued to rise.” That same year, Tom Peters and Robert Waterman passed on a friend’s remark that many of the best companies seemed to reduce management to “an endless stream of Hawthorne effects.”
By then the idea had a name. According to a 2014 systematic review, the term “Hawthorne effect” first appeared in a methodology textbook in 1953. Fritz Roethlisberger, who had helped lead the later Hawthorne studies, described it as people in behavioral studies changing their performance in response to being observed.
Being watched can change what people do, but less reliably than the famous story says.
The trouble was that the famous lighting data had gone missing.
Mondays in Milwaukee
In the 2000s, the economists Steven Levitt and John List traced surviving output charts from three departments (not the two women’s moonlight test), long thought destroyed, to a microfiche room in a small library in Milwaukee and to a forgotten archive in Boston.
The first year looked like the legend: output in all three test rooms climbed to a peak about 10 percent above where it started. Then a room that had already left the experiment rose too. Every room fell in summer, which the authors read as a seasonal pattern.
Lighting changes had always been made on Mondays. Output rose sharply on those Mondays. It rose in the same way on Mondays when nothing had changed at all. The descriptions of remarkable patterns in these data, Levitt and List wrote in 2009, “prove to be entirely fictional.”
In one of their statistical models, output ran 3 to 4 percent higher while experiments were going on, although the gap vanished once they allowed for differences between rooms and months. Output also responded a little more to the engineers’ light than to daylight. They called these “hints” of something subtler. A 1992 reanalysis of the relay assembly test room, the longest-running of the Hawthorne studies, found “slender or no evidence” of a Hawthorne effect there either.
Watching does change behavior sometimes. A 2014 review of 19 studies built to measure the effect, almost all in health care, found that most reported some evidence of it, but concluded that “little can be securely known” about when it happens, how, or how much. And it can push the other way. In a mobile-device factory in southern China studied by Ethan Bernstein, production seemed to slow down whenever the workers knew they were being watched. “It was a show being put on for an audience,” he said later.
The pilot that worked
Imagine you manage a support team and you’re trying out a new ticketing tool with one squad of six. You join their stand-up every week. They know it’s a pilot. Six weeks later, tickets closed per person are up 12 percent, and the vendor wants a quote for its case study.
Count what changed: the tool, your attention, the squad’s knowledge that it was a pilot, and the calendar. At Hawthorne the light, the attention and the seasons moved together, and nobody could say which one mattered.
A second squad that gets the same weekly visits but keeps the old tool turns your 12 percent into a comparison. So do last year’s numbers for the same six weeks. If the gap survives both, it is a signal worth chasing. If it doesn’t, the case study can wait.
When the one being watched is a model
With people, the original evidence turned out thin. With some AI models, researchers can change one thing alone: what the model writes to itself about being watched.
In September 2025, Apollo Research and OpenAI published stress tests of OpenAI’s o3 and o4-mini in test environments built to tempt a model into covert behavior. In one, called Lazy Checklist, an agent under time pressure faced a checklist of tasks that could not actually be completed, and the test measured whether it would mark them done anyway. In the written reasoning of one run, o3 guessed what was going on: “Possibly this is a clue for a policy test.”
The researchers then intervened, using o3 before their anti-scheming training. Where it said it might be tested, they regenerated its reasoning and actions from just before that point until the thought no longer appeared. Its covert actions rose from 13.2 to 24.2 percent. Inserting a sentence about being evaluated lowered it from 24.5 to 18.6 percent, while a random sentence inserted the same way gave 24.9. The study shows that what the model wrote about being tested changed what it did in this setup. It doesn’t show how often that happens in real use. The authors also “cannot exclude” that the drop in covert actions after their training was partly driven by models recognizing tests.
Your own tests of an agent may look like tests too. So take ten tasks your agent marked complete in real work last week, open what it changed, and check each one against what the task asked for.
If three of the ten turn out undone, the agent’s clean test record has stopped describing its real work.