Outlier
A value far from the rest of the data — which may be a broken measurement or the rare event you are trying to predict.
Think of It Like This
Like finding an elephant in a dataset of domestic house cats: it throws off the average weight calculation unless handled separately.
A height of 3.2 metres is a misplaced decimal point. A resting heart rate of 210 may be the event a model exists to catch. Detection can't tell them apart, so provenance decides the action rather than statistics.
The two common rules behave very differently. The IQR rule flags anything outside and is built from percentiles, so a single extreme value barely moves it. The 3-sigma rule uses the mean and standard deviation that the outlier itself inflated, so large outliers widen the threshold they're tested against and can hide entirely.
Deleting is the last option, not the first: investigate, correct, cap, transform, or switch to a model that tolerates the tail. Removing the tail to improve a metric usually removes the part you're graded on.