Ignoring how common something is when judging a test result. If one person in ten thousand has a condition and your test is 99 per cent accurate, most positive results will still be wrong, because there are so many more healthy people to generate false alarms. This arithmetic governs every screening and detection system.
Technical view
Neglect of the prior in Bayesian updating. For a rare positive class, precision is dominated by the false positive rate multiplied by the size of the negative class, regardless of how good sensitivity is. This is why accuracy is a useless measure under class imbalance, why alert fatigue is structural rather than a failure of attention, and why deployment thresholds must be set against the real prevalence rather than a balanced test set.
Related
Part of AI Terminology Index, section: Maths and Statistics.