Are AI detectors accurate?
What “accuracy” really means for an AI detector, the numbers we measured, and when you should not trust a result.
“99% accurate” is not one number
Accuracy hides a trade-off between two errors:
- False positives: human writing flagged as AI. This is the error that hurts people.
- False negatives: AI writing that slips through.
A detector can lower one only by raising the other. So look for precision (of the sentences flagged, how many are really AI), recall (of the AI sentences, how many were caught) and the false-positive rate, measured on data the detector has not seen.
TextFlag's measured accuracy
| Mode | Flagged sentences that are AI | AI sentences caught | Human sentences flagged by mistake |
|---|---|---|---|
| Standard (default) | 95.3% | 41.2% | 1.2% |
| Thorough | 70.5% | 80.5% | 20.5% |
Measured on 132 held-out documents (46,862 sentences). Full methodology →
For the document-level percentage: TextFlag is calibrated on 658 documents of known authorship, including AI text from GPT, Claude, Gemini and DeepSeek and 440 human papers written before 2019. Pre-2019 human papers average about 1% or less, and nine documents in ten land within 9 percentage points of their true AI share.
Where AI detectors struggle
- Short texts: under about 40 sentences, TextFlag's margin widens to about ±25 points.
- Lightly mixed documents: documents that are 20% AI tend to read about 13%.
- Some model families: pure Claude text averages about 92% rather than 100%.
- Paraphrased AI text and non-English text.
How to use an AI detector responsibly
Use the strictest mode when judging someone else, read the flagged passages yourself, and look for other evidence such as drafts and sources. A score alone should never decide a grade, a job or a publication.
Questions people also ask
Which AI detector is the most accurate?
Accuracy depends on the text and the mode, so compare detectors on published false-positive rates measured on held-out data. TextFlag publishes its precision, recall and false-positive rates for both modes.
Can AI detectors give false positives?
Yes, all of them. TextFlag's Standard mode flagged 1.2% of human sentences in testing; Thorough mode flagged 20.5% in exchange for catching more AI text.
Are free AI detectors accurate?
Price says little about accuracy; published, independently measurable numbers say more. TextFlag is free up to 10,000 words and publishes its methodology.
Related guides and tools
Check your own text for AI
Free for up to 10,000 words. Highlighted report and PDF included.