False Positive Rate
In AI detection, the false positive rate is the percentage of genuinely human-written texts that an AI detector incorrectly classifies as AI-generated.
Definition¶
The false positive rate (also called the false alarm rate) in AI text detection is the proportion of human-written documents that an AI detector incorrectly flags as AI-generated. A false positive rate of 5% means that, on average, 5 out of every 100 genuinely human-written texts will be incorrectly classified as machine-generated.
Why It Matters¶
False positives have serious practical consequences in academic and professional settings. A student falsely accused of submitting AI-generated work faces disciplinary action for something they didn't do. A journalist or author may have their credibility questioned based on a probabilistic score rather than evidence.
Published studies have found false positive rates ranging from 2% to over 20% depending on the detector and the type of writing being tested. The highest rates are typically found with:
- Dense academic writing — formal prose with complex sentences and technical vocabulary has naturally low perplexity
- ESL (English as a Second Language) writing — non-native writers often produce more formulaic, lower-perplexity text
- Instructional and list-format content — structured content with consistent formatting produces low burstiness
Acceptable vs. Unacceptable Thresholds¶
In academic integrity enforcement, most experts argue that the false positive rate of current AI detectors is too high to justify using detector output as definitive evidence of misconduct. A 5% false positive rate applied to a class of 200 students would produce 10 false accusations — even if no student used AI at all.
The consensus position among academic integrity researchers is that AI detector output should be treated as probabilistic evidence warranting further investigation (such as a conversation with the student), not as proof of a policy violation.