← Blog

The Interstitial Journal blog · Built on evidence, not hype

The Planning Fallacy: Why "Twenty Minutes" Is Never Twenty Minutes

You said the deck would take an hour. It took three. You are not surprised, exactly — this happens most weeks — and yet next Tuesday you will say "an hour" again, with the same quiet confidence. That gap between the estimate and the clock has a name, a long research history, and one stubborn property: knowing about it doesn't fix it.

By comun.ai · 13 July 2026 · ~6 min read

The finding: we underestimate our own timelines

Daniel Kahneman and Amos Tversky named the planning fallacy: the tendency to predict that a task will go roughly as well as it possibly could, producing completion estimates that skew optimistic.¹ Strong Roger Buehler, Dale Griffin and Michael Ross put it under a microscope, asking students to predict when they would finish real work — theses, assignments, everyday errands — and then recording when they actually finished. People finished later than they predicted, on average, and did so even when they were asked for a "worst-case" estimate that they were highly confident they'd beat.²

The part that should unsettle you is the next one. The same people could readily recall that their previous projects had run long. The memory was available; it simply didn't feed into the forecast.

Why it happens: the inside view

When you estimate a task, you do it by imagining the task — the steps, the order, the smooth path from here to done. Kahneman and Tversky called this the inside view: a forecast built from the specifics of this particular case, generally in its best-case form.¹ Moderate The inside view has no slot for the things that actually consume the hours — the interruption, the missing file, the login that expired, the colleague who "just needs two minutes."

The corrective is the outside view: ignore the story of this task and ask how long tasks like this one have historically taken. That's reference-class forecasting, carried into practice most visibly in project management, where Bent Flyvbjerg has argued for pricing new projects against the distribution of outcomes from a class of comparable past ones.³ Moderate It works — when you have the reference class. Most of us, for our own work, do not.

Two clues about the mechanism

It's specific to your own tasks. Buehler, Griffin and Ross also had observers predict when someone else would finish, and the observers were markedly less optimistic than the people doing the work.² Moderate Watching from outside, you are already taking the outside view — you don't have the seductive plan in your head, so you fall back on how these things usually go. The fallacy lives in the first person.

Unpacking the task raises the estimate. When people are made to break a task into its components before estimating, their totals tend to come out higher — and closer to reality. Justin Kruger and Matt Evans found this segmentation or unpacking effect across several everyday tasks.⁴ Moderate The lesson is mundane and useful: a lot of the missing time is in the sub-steps you never bothered to name.

The problem with using memory as your record

The obvious remedy — "just remember how long it took last time" — leans on a faculty that isn't built for the job. Memory for duration is reconstructive: we don't store a stopwatch reading, we rebuild an impression afterwards from what the episode felt like and what we believe about it.⁵ Moderate A morning that fragmented into six interruptions gets remembered as the one task you were nominally doing. Which means the reference class you're trying to consult barely exists: you have anecdotes, not data.

There's a compounding reason those hours vanish. Every switch between tasks leaves attention residue — part of your focus stays stuck on the previous task, degrading performance on the next one until it clears.⁶ Strong That drag is real time, but it doesn't attach itself to any task in your recollection. It just quietly disappears from the ledger. (We covered it in attention residue and the case for interstitial journaling.)

Where a timestamped log comes in

Here is the reasonable — and deliberately modest — application. If the outside view requires a reference class, and yours is missing because memory won't hold it, then the fix is to stop relying on memory. Write a line at the seam between tasks, with the time on it: what you were doing, and that you're leaving it. Do that for a few weeks and the log becomes the thing you never had — a record of what your work actually costs, with your own interruptions included.

You are not tracking time to optimise yourself. You are building the smallest possible evidence base about one person: you. Next time you say "twenty minutes," you have somewhere to look.

Where the evidence ends

Let's keep this calibrated. The planning fallacy itself is strong — a robust, repeatedly demonstrated tendency. The inside/outside-view explanation and reference-class forecasting are moderately supported; so are the unpacking effect, the self/other asymmetry, and the reconstructive nature of duration memory. Attention residue is well-supported.

What has not been established is the claim that "logging your own task durations will calibrate your estimates." Early That is a sensible inference assembled from well-studied parts — a personal reference class is exactly what the outside view asks for — but it has not been trialled as an intervention, and we won't invent a percentage to imply otherwise. Note too that the fallacy has survived plenty of debiasing attempts; a notebook is no guaranteed antidote to a bias that shrugs off being explained to you. Treat it as reflection, not treatment. More on our science page, and how your notes stay private on our privacy page.