I don't remember it exactly, and I'd like to read up on it again.
It is about experiment design and frequentist hypothesis testing. A sequence of N measurements are made, and then when evaluating the data, whether the data shows a significant effect depends on why the experiment was a terminated: because it was already predecided that N measurements need to be made, or because of some other criterion (I don't remember). The "paradox" is that the data is the same, but we have to know how the experiment would have continued if something else had happened. So a part of the plan that never got applied is required to decide whether the effect is significant.
I clearly remember that this has a detailed Wikipedia article, but I can't find it either through Google or Wikipedia search or browsing. Maybe it was not designated as a "paradox", but an effect, phenomenon, thought experiment or something else.