AI Detector Accuracy

AI writing detectors are being used to reject submissions, cancel book deals and fail students. The vendors advertise false-positive rates around 1 percent or lower. Independent tests paint a different picture: results range from detectors that really do hold near-zero false positives to tools that flag a century-old novel as machine-written. This page puts the vendor claims and the independent measurements side by side, with every figure linked to the source it was read from.

If a detector score has been used against you, read our companion guide: what to do when you are falsely accused of using AI.

GPTZero's own benchmarking page states an accuracy rate of 99 percent for AI versus human text and says it keeps its false positive rate at no more than 1 percent (GPTZero, published 2025-01-30, accessed 2026-07-20).Turnitin's public claim, as carried in the comparison below, is a document-level false-positive rate of less than 1 percent.

How to read the spread

Three patterns hold across the independent tests. First, the gap between the best and worst tools is enormous: in the Authors Guild test the same ten pre-2022 articles scored 0 percent AI on Pangram and up to 100 percent AI on Sidekicker.ai. Quoting "AI detectors are 99 percent accurate" without naming the tool is meaningless. Second, headline vendor numbers are best-case: document-level scores on long native-English text. Sentence-level scores and short passages run worse, which matters when an editor highlights one paragraph of your manuscript. Third, the errors are not evenly distributed. The Stanford study found seven detectors averaged a 61.22 percent false-positive rate on essays by non-native English speakers while scoring native-speaker essays almost perfectly. Plain, polished, conventional prose is the profile detectors misread most, and the Authors Guild reached the same conclusion about experienced human writers.

The published studies above test essays and articles. Nobody regularly tests creative writing, the category this site's readers actually submit. So we are committing our own instrument in advance, before any tool is scored:

• Corpus: 20 excerpts of 300 to 500 words each (15 prose fiction, 5 poetry), all from works first published between 1813 and 1925 and drawn verbatim from Project Gutenberg files. Every excerpt predates generative AI by decades, so any AI flag is a false positive by construction. The full corpus, with per-work source URLs, access dates and character offsets, is committed in the site repository as detectorCorpus.json, and the extraction is scripted so anyone can reproduce it.
• Method: each quarter, every excerpt is run once through each consumer-facing detector on its default settings, on the same day. We record the per-excerpt AI score verbatim, publish the full grid rather than a summary number, and never re-run to get a nicer result.
• Reporting: results appear on this page as a dated study block alongside the third-party evidence, never blended with it.

The corpus leans deliberately on early-20th-century plain prose (Stein, Anderson, Cather, Lewis, Fitzgerald): the published evidence says clean, uncluttered prose is exactly what detectors misread. The first run has not happened yet. Until it does, this page carries only the third-party results above.

Where this matters for your submissions

Magazines and contests are writing AI rules directly into their guidelines, and publishers are acting on detector scores. See which markets ban, allow or require disclosure of AI at our AI submission policies tracker, and if you self-publish on Amazon, the KDP AI disclosure wizard walks the generated-versus-assisted line. If you have already been accused, start with the defense guide.