What is the planning fallacy?
The planning fallacy is the systematic tendency of people to underestimate the time (and cost, and risk) of their own future work, even when they know perfectly well that similar work has overrun in the past. The term was coined by Daniel Kahneman and Amos Tversky in 1979, and its most uncomfortable finding has replicated for decades: the bias survives experience, incentives, and even explicit education about the bias itself. Students taught the planning fallacy in the morning underestimate their own assignments in the afternoon.
It is a bias, not noise: the errors point one way. That distinction is what makes it correctable.
The mechanism: inside view versus outside view#
Kahneman and Tversky's explanation has held up. Asked to estimate a task, people take the inside view: they mentally simulate this particular task going well, step by step, and read the duration off the simulation. The simulation rarely includes the interruptions, the rework, the dependency that arrives late, because none of those are part of the plan, only of reality.
The outside view ignores the particulars and asks a colder question: how long did the last ten things like this take, regardless of why? For prediction, the outside view wins, consistently, because it includes all the things that go wrong without needing to name them. Humans default to the inside view for their own work; the entire correction industry is ways of forcing the outside view back in.
| Inside view | Outside view | |
|---|---|---|
| Asks | "How will this task unfold?" | "How did the reference class turn out?" |
| Includes surprises | Only imagined ones | All historical ones, unnamed |
| Typical output | Best case wearing a typical-case costume | A base rate with a spread |
What does not fix it, and what does#
Trying harder does not fix it (the bias is in the simulation, not the effort). Padding does not fix it either: padding moves the centre by a guessed amount while still hiding the spread, and chronic padders just train their colleagues to discount them.
The corrections that work are all structural versions of the outside view:
- Reference-class forecasting. Estimate from the measured history of similar work, then adjust for specifics, not the other way round.
- Separate bias from variance. Track the ratio of actuals to estimates (your personal exchange rate) and apply it as a multiplier; express the remaining honest randomness as a range on each estimate.
- Let the system do the remembering. Software holds reference classes with perfect patience. Topolog records completion times as you work and recalibrates a per-area multiplier once there is real evidence, so the forecast reflects how you actually work rather than how you hoped to (the systems argument in full).
The reframe worth keeping: the planning fallacy is not a character flaw to overcome but a known property of the human estimator, like a sensor with a documented offset. You do not fix a sensor by exhorting it. You calibrate it, downstream, continuously, and then its readings become useful exactly as they are.