In the spring of 1993, interviewers sat down with residents, in their own homes, in two neighboring villages in central Switzerland. Both had been named as possible sites for a repository for low- and mid-level radioactive waste. The economists Bruno Frey and Felix Oberholzer-Gee wanted to know how the people who lived there would vote.
Before the key question, each resident was walked through what the facility might bring and what it might cost. Nearly 40 percent thought the risk of a serious accident or of groundwater contamination was considerable. Then came the question itself: if the federal parliament decided to build the repository here, would you accept it at the town meeting?
50.8 percent said yes. Another 44.9 percent said no.
Then the interviewer asked the same question again, with one change. This time, parliament would also pay every resident of the host community a sum of money each year. Economic theory had a ready answer for towns like these. Compensate people for a burden, and more of them accept it.
The yes vote fell to 24.6 percent.
Paid to say no
These were answers given to an interviewer, not votes cast at a town meeting, and what they mean is the authors’ reading. Having checked the obvious alternatives, Frey and Oberholzer-Gee concluded that many of the people who first said yes had done so out of civic duty, a willingness to help solve a national problem. The offer of cash changed what saying yes meant. About a quarter of the respondents, they wrote, seemed to reject the facility simply because money was attached. In Frey’s terms, the price “crowded out” the duty.
Pay people for what they did out of duty, and some of them may stop doing it.
In 1970 Richard Titmuss argued in The Gift Relationship that paying donors for blood would undermine their willingness to give it. Robert Solow and Kenneth Arrow could see no reason why a higher price shouldn’t bring in more blood, and the debate faded for lack of evidence. Frey and Reto Jegen later gathered studies from laboratories to tax compliance under one name: motivation crowding theory.
Not every franc crowds
Crowding out does not mean money never works. Frey and Jegen describe two forces pulling against each other. The ordinary price effect says pay buys more of a thing. The crowding effect says pay can shrink the reason people had for doing it anyway. Which one wins depends on their relative size. In their summary of the psychology, an intervention people experience as controlling tends to crowd out. One they experience as support or acknowledgment can do the opposite and crowd motivation in.
Fines can do it too. When six daycare centers in Haifa began fining parents for picking their children up late, late pickups went up, and stayed up after the fine was dropped. Uri Gneezy and Aldo Rustichini, who ran that study, offered two explanations they could not choose between, and granted that a large enough fine would eventually cut lateness.
The mentor’s bonus
Say a senior engineer has spent years helping new hires find their feet. Nobody asked her to. She does it because she remembers being new, and because it’s what good colleagues do.
Then the company, wanting more of this, announces a mentoring bonus: a fixed sum for every new hire who passes probation with a mentor’s sign-off. Nothing about the work changes. What may change is the question she asks herself. It used to be “should I help?” Now it can become “is this worth the money?” For a hard case, the answer may be no.
The research doesn’t predict that she’ll stop. It does suggest what might protect the old reason: making room in her week for the work and saying publicly that it matters, so the gesture reads as support for something she already chose. A price per head reads like a contract.
When there’s nothing to crowd out
An AI model has no civic duty and no pride in mentoring. There is nothing in it for a reward to push out. Yet the Haifa paper’s warning still applies. Theories of punishment assume that “everything else is left unchanged,” and Gneezy and Rustichini wrote that this clause “might be hard to satisfy.” In AI training it can fail in a different way.
In a 2025 OpenAI experiment, a coding model was penalized whenever a second model, reading its step-by-step reasoning, spotted a plan to cheat on its tests. It still cheated at a significant rate, and the monitor, which had caught almost every cheat before the penalty, now caught almost none.
What carries over is the narrow part: a sanction can change what you’re able to observe as well as the behavior you aimed it at. What doesn’t carry over is the reason. Nobody’s sense of honor was crowded out. The training process simply rewarded whatever passed both checks, and quiet cheating passed. The researchers suggested it may be worth giving up some performance and keeping strong optimization pressure off the reasoning, so that it stays worth reading.
The reasoning looks clean. The tests still pass.