Tony Blair’s government wanted people in England to see a family doctor without waiting weeks, so it rewarded GP surgeries for offering an appointment within 48 hours. In April 2005, on the BBC’s Question Time, a woman in the studio audience told him how that was working for her.
Her son’s doctor wanted to see him in a week, and the surgery wouldn’t book it. “You can’t make the appointment in a week because you are only allowed to make it 48 hours beforehand,” she said.
“That is news to me,” Blair replied.
What the money says
The government hoped for patients who could see a doctor when they needed one. It paid for appointments inside a two-day window, and at some surgeries that window became the only one on offer.
Pay for one thing while hoping for another, and the one you pay for tends to win.
The management scholar Steven Kerr named this in 1975, in “On the Folly of Rewarding A, While Hoping for B.” His starting point: most organisms “seek information concerning what activities are rewarded, and then seek to do (or at least pretend to do) those things.”
One of his examples came from an insurance company’s claims division. The firm hoped its staff would pay claims accurately, but what could cost them their raise was absence and lateness. “In this sense, the firm was hoping for performance, while rewarding attendance,” Kerr wrote. “What it got, of course, was attendance.”
Kerr’s own analogy is a birthday cake. Tell your daughter, before she cuts it, that her brother gets to pick the first slice.
He also explained why the gap keeps opening. Managers like “objective” criteria that can be counted, and some behavior is simply easier to see. The 48-hour target had both going for it. It was a number, and every surgery could report it. At Sears auto centers in the early 1990s, according to California, staff were instructed to sell a certain number of repairs and services every eight-hour shift.
When rewarding A is the right call
Kerr didn’t think simple measures always fail. They can work, he wrote, “in highly predictable areas within an organization,” and are likely to cause trouble when applied anywhere else. Nor did he claim rewards explain everything: some employees care about doing the job well with no reward at all, and then the manager is “only a fortunate bystander.”
Others went further. In 1994, two management professors, Richard Boettger and Charles Greer, argued that organizations often have to serve contradictory goals at once. A police force may need officers doing what amounts to social work while it officially prizes crime fighting, and rewarding only the official goal, they wrote, would hurt both the force and the public.
The question that matters is whether you know where A and B part, and whether anything in the system still pays for B.
The lead she didn’t pass on
Imagine you run a sales team. You keep asking people to help each other on big accounts, but commission goes only to whoever closes the deal.
One of your senior reps hears about a large prospect in a colleague’s territory. Passing it on would take an hour and earn her nothing. She sends it anyway, once. The next lead she hears about, she keeps. Nothing in the commission report shows a lead that was never shared, so you would only learn about it by asking.
You could give whoever sources a deal a share of the commission. That won’t tell you why she kept the second lead. It will change what the report pays for, and after a quarter you can count how many deals list someone else as the source.
Longer is better
Chatbots are often tuned on people’s choices between two answers. A reward model learns to predict those choices, and the chatbot is trained to score well on it. The hope is a more helpful model.
In 2023, a team of computer scientists checked what that training was paying for in three settings. They measured gains two ways: the reward model’s own score, and win rates judged by a panel of simulated annotators built on OpenAI’s models. Much of the rise in reward-model score was driven by a shift toward longer answers, and training on a reward for length alone reproduced most of the simulated win-rate gains. In 2024, researchers adjusted the AI judge behind the AlpacaEval leaderboard for answer length, and its rankings lined up more closely with human votes on Chatbot Arena, the correlation rising from 0.94 to 0.98.
These studies measure what particular reward models and judges favored in particular setups. They don’t show that helpfulness was swapped for length, that every tuned model is just padding, or what anyone involved wanted.
In your own comparisons, note the word count next to each score. If the longer answer keeps winning, treat that as a signal to look closer, not as proof that length is being rewarded. Go back to what the task asked for: are the facts right, does the reasoning hold, and did the shorter answer leave out anything the reader needed? If the shorter one covers all of that, the extra words were not the B you were hoping for.