also searched as: is gptzero accurate · is gpt zero accurate · is gptzero ai detector accurate · how accurate is gptzero ai detector · does gptzero work · is gptzero a good ai detector · gptzero false positive
GPTZero is accurate on unedited AI text and rarely flags human writing, but a humanizer or heavy rewrite can cut its catch rate in half. Independent tests put its false positives near 1%, higher on short text. Treat a high score as a reason to talk, not as proof.
Trust it for a first screen of long, unedited text, where it misses few AI passages and flags about 1% of human ones.
Double-check it when the text is short, was paraphrased or humanized, or the writer's first language is not English.
Don't rely on it if the score alone would decide a grade, a job or a misconduct case.
sources checked 2026-10-13 · by dan scutaru
| who tested it | what they found |
|---|---|
| Chicago Booth / NBER study, 2025 | about 0.7% of human text flagged 1,992 human passages in six genres matched with GPT-4.1, Claude and Gemini versions; it missed 5% or less of medium and long AI passages and flagged 2.4% of short human reviews. source |
| Chicago Booth / NBER, humanized text, 2025 | missed about half or more Every AI passage rewritten by StealthGPT; GPTZero missed around 50% or more across most genres, and disputes which API field the authors read. source |
| RAID benchmark, UPenn and others, 2024 | 66.5% caught at a 5% false-positive cap Queried January 2024 across 11 generators and 8 domains; under 10% on sampled open-source base-model text, and only 2.5 points lost to paraphrasing. source |
| Bloomberg Businessweek test, Oct 2024 | 1% to 2% of human essays flagged 500 Texas A&M application essays from summer 2022, before ChatGPT, run through GPTZero and Copyleaks; reported for the two tools together. source |
| Stanford (Liang et al.), March 2023 | 47 of 91 non-native essays flagged TOEFL essays by non-native writers, against 0 of 88 US eighth-grade essays; our count from the study's published scores at a 0.5 cutoff, March 2023 version. source |
| GPTZero's own claim | 99% accuracy, about 1% false positives Claims 99% accuracy, 95.7% caught at 1% false positives on the RAID leaderboard, and 0.05% false positives in its own rerun of the Booth data. source |
Very, as long as nobody touched the text. In the Chicago Booth study it missed 5% or less of medium and long AI passages, and 7% of short ones. The University of Pennsylvania's RAID benchmark was harsher in early 2024: across 11 generators it caught 66.5% when held to a 5% false-positive cap, because it barely recognised some open-source models. GPTZero has retrained since; its newer, higher numbers are its own.
Mostly, with one big exception. On long human text the Booth study measured false positives of 1% or less, and Bloomberg found 1% to 2% of pre-ChatGPT admissions essays flagged. The exception is non-native writers: in March 2023 GPTZero flagged 47 of 91 TOEFL essays. GPTZero said in December 2023 that a retrained model hit a 1% target on that test, but we found no independent recheck. Whether any detector is accurate enough is covered on our page about AI detectors.
It depends on what did the rewriting. Simple paraphrasing barely hurt it in RAID, whose authors called it unusually robust to attacks. Dedicated humanizers are another matter: after StealthGPT rewrote the AI passages in the Booth study, GPTZero missed around half or more of them. GPTZero says the authors read the wrong field of its API, but its published rerun does not cover the humanized set.
Yes. It is a real detector, released in January 2023 and acquired in June 2026 by Superhuman, the company formerly called Grammarly. Legit is not proof, though: its own site says results should not be the final verdict.
Not in the newest independent test. The Chicago Booth study ranked Pangram first and put GPTZero in a second tier with Originality.ai, ahead of it on false positives.
GPTZero, on the one independent benchmark we found that tested both. In RAID it flagged essentially no human text at default settings, against 1.42% for ZeroGPT, with similar catch rates.
Yes. Independent tests put it around 1% on long human text and up to 2.4% on short reviews, some with near-100% certainty in Bloomberg's test.