Douglas Hofstadter put it in Gödel, Escher, Bach in 1979, and the recursive clause is what makes it more than a joke: it always takes longer than you expect, even when you take into account Hofstadter's law.

Knowing about the bias does not correct it, which is unusual and worth understanding.

Why the buffer does not save you

An estimate is built by imagining the work: these steps, this long each, added up. The picture is of the project going as planned, because that is the only version anyone can picture in detail.

What actually consumes the time is the set of things not in the picture — the dependency that was not ready, the question nobody could answer for three days, the rework after a decision changed. None of these are in the plan because they are not knowable in advance, and there are always some.

Padding the estimate does not help much, because the padding is a percentage of a number generated by the same flawed method. Twenty percent more of an imagined smooth run is still an imagined smooth run.

What works instead

Use the outside view. How long did the last five comparable projects take? That number already contains all the interruptions nobody can predict individually, and it is almost always longer than the bottom-up estimate. When the two disagree, the outside view wins. This is base rate neglect applied to your own calendar.

Estimate in ranges and commit to the top. A range forces the question of what the bad case looks like, which the single number hides.

Decompose only to where you have history. Breaking work into smaller pieces helps when you have measured pieces of that size before, and adds false precision when you have not.

Track your own multiplier. The ratio of actual to estimated is stable per person and per kind of work. Measuring it once gives you a correction factor with evidence behind it, and it is usually larger than anybody wants to admit.