Slice a dataset into subgroup after subgroup until chance hands you p < 0.05, then watch a fresh sample decide whether your finding was real or a fluke.
The treatment has no real effect, so any significance you find is noise. Filter the data into subgroups and re-run the test as many times as it takes to stumble onto a p-value below 0.05, then submit your 'finding' to the referee. The twist: a replication check draws a fresh independent sample and re-tests your exact claim. Real effects survive; p-hacks vanish, and that gap is the whole lesson.
Cherry Picker is about p-hacking, multiple comparisons and why replication matters, the integrity of statistics.
Filtering into subgroups and re-running the test until something turns up significant is data dredging — p-hacking: performing many tests and reporting only those that come back significant.
The original design is that the treatment has no real effect, so every result you find is noise by construction.
That is the only honest way to teach this. Told in the abstract, p-hacking sounds like cheating that a careful person would never do; performing it on data guaranteed to be empty shows how easily significance appears, and this is the multiple comparisons problem rather than a character flaw.
Cherry Picker is a statistics game about p-hacking, multiple comparisons and why replication matters. It's an original free browser game built for PlayPendium.