About a year into writing a high school textbook on judgment and decision making, Daniel Kahneman tried an exercise on his own team. Each member wrote down, privately, how long it would take to hand Israel’s Ministry of Education a finished draft. He wrote the answers on the blackboard. They clustered around two years. The lowest was a year and a half, the highest two and a half.
Then he turned to Seymour Fox, the team’s curriculum expert, and asked about other teams that had built a curriculum from scratch. How much longer had they needed from this point?
Fox fell silent. Then he admitted something he had never realized before: not all the teams at their stage ever finished. About 40 percent, he estimated, never did. Kahneman asked how long the rest had taken. “I cannot think of any group that finished in less than seven years,” Fox said, “nor any that took more than ten.”
Fox’s own estimate was on the blackboard too, inside the two-year cluster.
The plan in front of you
We forecast from our own plan and forget what happened to everyone else who tried.
Kahneman and Amos Tversky gave the error its name in a 1977 technical report, published in 1979. “Scientists and writers, for example, are notoriously prone to underestimate the time required to complete a project,” they wrote, “even when they have considerable experience of past failures to live up to planned schedules.” They called it the planning fallacy, and noted that it often shows up even when underestimating is penalized.
Their explanation was arithmetic. A building, they wrote, can only be finished on time if there are no delays in deliveries, no strikes, no unusual weather. Each of those is unlikely. The chance that at least one of them happens may be substantial, and intuition doesn’t add that up. A plan is a list of the steps you can picture, and the delays that sink projects are the ones nobody pictured. Kahneman’s team later listed theirs: divorces, illnesses, crises of coordination with bureaucracies.
Fox’s two answers show the two ways to forecast. Asked about his own team, he looked at the work in front of him and said two years. Asked about other teams, he looked at what had happened to them. Kahneman and Tversky later called these the inside view and the outside view. The book was finished eight years after that discussion.
The same pattern turns up with money. Zillow set itself a guardrail of 2 percentage points either side of breaking even on each home it bought to resell. Between the second quarter of 2021 and its outlook for the fourth, the swing was about 12.
Fifty-five days
Kahneman himself warned against making too much of one story. His team’s outcome matching Fox’s numbers so closely, he wrote, was surely a fluke. The firmer evidence came from experiments.
In a 1994 study, Roger Buehler, Dale Griffin and Michael Ross asked 37 psychology students at the University of Waterloo, in the final semester of their honors theses, when they’d hand them in. Among the 33 who finished, the average guess was 33.9 days and the average actual time 55.5. The other four still hadn’t finished when the researchers stopped tracking, two semesters later. Fewer than a third finished by their best guess. They were also asked for a date “if everything went as poorly as it possibly could.” Fewer than half met even that one.
The same paper found where the bias weakens. People asked to predict someone else’s finish leaned the other way: they overestimated others’ completion times and paid more attention to past experience. And in one study the optimism disappeared when participants were told to connect their past experiences to the new prediction. The outside view comes with a condition. Kahneman’s argument for it holds “if the reference class is properly chosen,” and choosing it is a judgment call of its own.
When the date is yours to give
Imagine you’re asked when your team’s move to a new billing system will be done. You walk through the plan: export, mapping, testing, cutover. It adds up to six weeks, and it feels honest.
Before you send it, you open the tracker and find the last three migrations your team ran. Each was planned at about six weeks. Each took between ten and fourteen, and none of the delays was in its plan.
Now you can give two numbers: the six weeks your plan says, and the ten to fourteen your team’s record says. Whoever has to staff the launch gets to choose which one to bet on.
The speed-up forecast
Adopting an AI tool is a forecast too: this will go faster. In early 2025, the research group METR put that forecast to a test. Sixteen experienced open-source developers listed 246 real tasks from their own projects, and each task was randomly assigned to allow AI tools or not.
Before starting, the developers expected AI to cut their time by 24 percent. With AI allowed, they took 19 percent longer. Afterwards, they still believed AI had sped them up by 20 percent.
The study is a snapshot of early-2025 tools and experienced developers in code they knew well, not of developers in general today. In February 2026 METR wrote that it believed developers were likely more sped up by newer tools, but that selection effects in its follow-up experiment made its data only very weak evidence of how much. The original randomized trial was not withdrawn.
Your own reference class is easy to start. For the next ten tasks, note whether you’ll use AI and how long you expect the task to take, then log the actual time. After ten, put the column of guesses next to the column of clock times. That shows how well you estimate the way you actually worked. It can’t tell you how long the same task would have taken the other way, or how much the AI itself sped you up.