A false positive occurs when an automated system (classifier, detection tool, algorithm) incorrectly labels a legitimate case as problematic. In information integrity research, false positives have serious consequences: labeling genuine humans as bots can harm reputation and limit legitimate speech; tagging authentic news as misinformation can chill reporting; flagging legitimate accounts as coordinated inauthentic behavior can suppress political organizing.
Problem magnitude
In imbalanced classification problems (where negative cases vastly outnumber positive ones — e.g., 85% humans vs. 15% bots on Twitter), even a low false-positive rate on the model's training set can produce unacceptable false-positive rates in practice. A classifier achieving 95% overall accuracy can still mislabel most of a minority class.
Key papers
- 2020-rauchfleisch-bot-detection-false-positives — Demonstrates that Botometer, despite good diagnostic ability (ROC-AUC 0.85), produces 41–76% false-positive rates when applied to realistic Twitter populations due to class imbalance; argues that ROC-AUC alone is misleading for imbalanced data and that precision-recall curves are essential for evaluating detection tools in practice
Connections
- bot-detection — false positives/negatives trade-off in bot detection
- measurement-and-evaluation — statistical foundations for understanding error rates