GPTZero

GPTZero flagged your essay. Here is what its numbers are.

Two outputs, one guess. GPTZero gives a document-level judgment about how likely the text is AI-generated, and highlights the individual sentences that most drive that judgment. Since autumn 2023 it has worked with a deep-learning model (the older perplexity-and-burstiness description still circulates in blogs, including terms GPTZero itself has moved past). Either way, the output is a probability about patterns — not a record of who wrote the words.

Its accuracy claims are its own. GPTZero publishes a 99% accuracy rate and a false-positive rate of “no more than 1%” on its benchmarking page — vendor-run numbers, on vendor-chosen samples. Independent testing tells a less tidy story: Bloomberg Businessweek ran GPTZero and Copyleaks over 500 college application essays written before ChatGPT existed — guaranteed human — and the services falsely flagged 1–2% of them, in some cases with near-100% certainty. Stanford's study found the detector class of that era flagged 61% of non-native speakers' essays; GPTZero says it has since worked to de-bias for second-language writers — a vendor claim worth quoting back if it is now being used against one.

GPTZero's own warning, from its FAQ: “these results should not be used to punish students,” and its educator guidance says a detector result should “never be the sole basis for action.” It also concedes its training data is mostly English prose by adults, and that it can flag “highly procedural text” — lab reports, standard forms, careful formulaic writing — as machine-made.

How to answer a GPTZero case: ask for the sentence-level report, then answer the highlighted sentences one at a time — where each came from in your drafts, what your notes said before the sentence existed. Sentence-level accusations invite sentence-level provenance, and your drafts have it. The checker builds exactly that item into your evidence list when you choose GPTZero.

Sources: GPTZero FAQ and limitations pages and AI accuracy benchmarking page (gptzero.me), checked October 19, 2026; Bloomberg Businessweek detector test, October 2024; Liang et al., Patterns, 2023 (Stanford HAI summary).