False positive rate.

False positive rate

False positive rate (FPR) = false positives / (false positives + true negatives). Or: out of every 100 human essays, how many get flagged.

This is the number that matters most for classroom deployment. An FPR of 1% means one in every 100 honest essays is wrongly flagged, which sounds low until you remember that a single teacher grading 80 essays will see one false flag every few weeks. We publish quarterly FPR on /transparency, including stratified FPR by known risk group.

Translating FPR to classroom impact

An FPR of 1% sounds harmless until you do the multiplication. A teacher grading 80 essays per assignment, across four assignments per semester, sees 320 essays. At 1% FPR, that's three wrongly flagged honest essays per semester per teacher. Across a 200-faculty department, that's 600 false flags per semester, most of them resolved, some of them escalated, some of them harming students. The math is why FPR is the metric that matters most for institutional procurement.

Stratified FPR, the number to ask for

Aggregate FPR conceals the per-population numbers. A detector with overall 1% FPR might have 3% on non-native English writers and 0.5% on native ones. The non-native number is what an ESL student feels. We publish stratified FPR quarterly on /transparency, including the populations where our number is worse than we want it to be.

How false positive rate interacts with related metrics

False positive rate is mathematically independent of precision, a distinction that often confuses practitioners. While FPR is calculated as FP / (FP + TN) and depends only on the population of true negatives, precision is FP / (FP + TP) and depends on the population of predicted positives. A classifier can exhibit a low false positive rate of 2% yet still have poor precision if the base rate of actual positives is very low. In AI essay detection, where genuine human essays vastly outnumber AI-generated submissions in many educational contexts, a 2% FPR can translate to hundreds of wrongly flagged students even when precision appears acceptable.

The relationship between false positive rate and specificity is direct and inverse: specificity equals 1 minus FPR. This means a system with 98% specificity has a 2% false positive rate by definition. Sensitivity (true positive rate) operates independently on the positive class, meaning developers can improve sensitivity without affecting FPR if they adjust decision boundaries carefully. However, in practice, lowering classification thresholds to catch more AI-generated text typically increases both true positives and false positives simultaneously, creating a tradeoff that ROC curves visualize by plotting TPR against FPR at various thresholds.

Practical implications for institutions and educators

Educational institutions must translate false positive rates into expected student counts to assess real-world risk. A university processing 10,000 essays with a detector exhibiting a 3% FPR will statistically flag approximately 300 innocent students, assuming all submissions are human-written. This volume creates administrative burden, erodes student trust, and exposes institutions to appeals and potential legal challenge. Administrators often focus on headline accuracy figures (such as 97% overall accuracy) without recognizing that FPR specifically quantifies harm to students who followed academic integrity policies, making it the more relevant metric for policy decisions.

Responsible implementation requires institutions to establish secondary review processes that account for expected false positive volumes. Some universities adopt a two-stage verification system where automated flags trigger human review rather than direct accusations, effectively treating the detector's output as a screening tool rather than definitive evidence. Others set threshold policies that accept higher false negative rates (missing some AI text) to achieve false positive rates below 1%, prioritizing the protection of honest students. The choice reflects institutional values, but both approaches require explicit acknowledgment that any automated system operating at scale will misclassify some number of legitimate submissions regardless of its reported performance metrics.