A classifier may find a correlation that works in familiar examples but fails when the background, source or measurement process changes. A high score can hide reliance on that correlation.
Researchers can look for shortcuts by changing suspected cues while holding other features stable. Such tests help explain whether a model's apparent success transfers to different conditions.
ELI5
Shortcut learning is like getting answers right for a reason that stops working when the question changes.
For example, imagine recognizing a dog only because the picture has grass. A dog photographed indoors could then confuse the model.
Is a shortcut useful on every dataset?
It can predict labels on data that preserve the same correlation. Reliability may drop if the relationship changes in another setting.
How can an evaluation reveal a shortcut?
Include cases where the suspected cue and correct label vary independently, then check how predictions change.

