Carroll and Rosson named this in 1987 after watching people learn software. Nobody read the manual. They started pressing things, got something that half worked, and then kept the slow method they had invented long after a faster one had been sitting two pages into a document they never opened.
The paradox is not that the behaviour is foolish. It is that the behaviour is correct. Reading costs time now, against a saving that is uncertain, undated and usually written for a slightly different version of the thing in front of you. Pressing costs nothing to begin and pays immediately, in feedback you can see. Offered those two, most people take the second, and most of the time they are right to.
Below is a task with a rule in it you cannot see from the outside. Four dials, a target, and a button holding 79 words that say exactly how the dials behave and which order gets you there in the fewest presses. The dials are live from the moment they appear. Nothing stops you pressing.
Two numbers come out at the end. How many presses you spent, and the shortest route from the start you were dealt, which the page knows exactly because the arithmetic is fixed. Whether you opened the panel is recorded as well, because it is the one thing here that is a single bit and not an interpretation.
What the second number is not
One run has one condition. You opened the panel or you did not, and there is no second version of you who did the other thing, so nothing here can tell you that reading would have been cheaper for you. The presses above the minimum are a fact about your run. They are not evidence about reading, and this page is not going to convert one into the other.
The 79 words are about 19 seconds at an ordinary pace, and that is a real cost paid up front against a saving nobody has promised you. Someone who reads all of it still has to work out where to start. Someone who reads none of it can see the coupling on the second press and be finished before the reader has reached the end of the paragraph.
What it costs you
Instructions are not a control. If something only behaves correctly when a person has read a sentence first, it does not behave correctly. The sentence is a hope, and it is the first thing skipped by people in a hurry, which is everyone.
The first method that works is the one that gets kept. Not the best one. The first one. Efficiency arrives long after the habit has set, if it arrives at all, and a feature added later has to beat a habit rather than an empty slot.
Put the rule where the hand is. A note attached to the dial is read at the moment it is needed. The same note two paragraphs above the dial is read by the people who were going to manage anyway.
The person exploring is not being careless. They are running a cheaper experiment than the one you wrote, and they get an answer in one press. Design that admits this — undo, forgiving defaults, states that explain themselves — outperforms any amount of front-loaded instruction.
I wrote the panel three times trying to make it short enough that skipping it would look unreasonable. It is 79 words now and I still skipped it on the last check, because I already knew the rule and wanted to see the dials move.