One afternoon in the spring of 2006, a math teacher at Parks Middle School in Atlanta unlocked the room where the state tests were kept. The booklets were wrapped in cellophane. Scissors would leave marks, so he walked to the corner store and bought a razor blade.
Back in the room, by his own account to The New Yorker, he slit the cellophane, warmed the blade with a lighter and worked it under the seal of a booklet. He photocopied three sections, then put on gloves and melted the cellophane shut again, so it looked as if nobody had touched it.
Unless enough students passed, the state could shut the school down. At the end of testing week, he and another teacher spent about an hour erasing wrong answers and bubbling in right ones.
A month later the scores came back, and the school laid out ice cream, pizza and hot wings in the hallway. For the first time since No Child Left Behind became law, Parks had met its annual goals. The share of eighth graders passing rose 31 points in reading and 62 in math. “It was like our World Series, our Olympics,” one student said.
The scores also set a new floor. Next spring, Parks would have to beat them.
The number, and the thing under it
Goodhart’s Law says that once a measure becomes a target, it stops being a good measure. In 1976, in a paper on how to evaluate social programs, Donald T. Campbell wrote down two “pessimistic laws,” packed into one sentence: “The more any quantitative social indicator is used for social decision-making, the more subject it will be to corruption pressures and the more apt it will be to distort and corrupt the social processes it is intended to monitor.”
Campbell leans hardest on the second half. The pressure reaches past the number into the work it was supposed to watch. About schools, he was blunt: when test scores become the goal of teaching, “they both lose their value as indicators of educational status and distort the educational process in undesirable ways.”
A thermometer you are graded on gets fiddled. Grade people on it for years, and they start heating the room to suit it.
That is what Georgia’s special investigators found when they reported in 2011. They confirmed cheating in 44 of the 56 Atlanta schools they examined and named 178 educators, 82 of whom confessed. The superintendent, Beverly Hall, ran the district on yearly test targets: “No exceptions. No excuses.” She replaced 90 percent of principals during her tenure.
At Parks, the changed answers were only the start. The principal wrote each teacher’s target in marker on the floor at the classroom door, where students could see it, and told teachers, “You need to teach to the test.” Lessons ran on deadlines. “We had two weeks to teach percentages,” the math teacher recalled, and a teacher still on percentages in week three was told he didn’t teach well enough. Attendance began to be taken later in the day. Eventually, he said, teachers stopped marking absences at all.
Visitors came to study the turnaround, so teachers built “standard-based mastery folders” to explain it. The math teacher wrote his master’s thesis on the method. “But we only put it in place to hide the fact that we were cheating,” he said.
Even the teachers lost their feedback. His wife taught language arts at Parks, and one year her students had the highest scores in the building. He never told her that other teachers had been correcting her students’ answers.
Second place on the leaderboard
In April 2025, Meta’s new Llama 4 Maverick model shot to second place on LMArena, a site where people compare two chatbots’ answers and vote for the better one.
Then researchers read the fine print. The version Meta had entered was an “experimental chat version” that was “optimized for conversationality,” not the model it released. LMArena said Meta’s reading of its policy “did not match what we expect from model providers” and updated its leaderboard policies. A separate paper that month counted 27 private variants Meta had tested on the arena before the Llama 4 release, and argued that practices like this lead to “overfitting to Arena-specific dynamics rather than general model quality.”
That is Campbell’s second half again. Developers pick models by the leaderboard, and the leaderboard reached back into how models are built. Simon Willison, a developer who had been impressed by the ranking, said afterward: “I can’t even use the model that they got a high score on.”
You can check this one yourself. Before you choose a model by its rank, look up the exact model name on the leaderboard, confirm it matches the one your provider serves, and run twenty of your own real tasks through that model.
The file at the bottom of the pile
Imagine you run a job placement office. Last spring the city began ranking your caseworkers by people placed. Placements are up. You sort the open files by how long each person has waited.
The top of the list is full of people with the hardest cases: no recent work, no car, a record. Campbell described this in an employment office: rate staff on placements, and they concentrate on the easiest cases. Nobody broke a rule. The work simply drifted toward whoever counted fastest.
Campbell also noticed something about trust. In Chicago, he wrote, people trusted the census more than the vote count, even though the census had fewer safeguards. Votes decided jobs, money and power. The census, until recently, had decided almost nothing.
So you keep a quiet number. Once a month, someone who isn’t ranked on anything reads ten files from the bottom of the pile and notes how long each has waited. The number never appears in anyone’s review.
The first month it comes back, three of the ten have been waiting more than half a year.