Is GPTZero accurate compared with Turnitin, Copyleaks and Originality.ai?

Vendor claims and independent tests are different things

Each of the four companies publishes an accuracy figure measured on data the company chose. GPTZero's technology page says it keeps its false positive rate at no more than 1% and reports 96.5% accuracy on mixed human and AI documents. Turnitin's AI writing FAQ commits to a document-level false positive rate under 1% for documents with more than 20% AI writing, validated against 700,000 pre-ChatGPT papers. Copyleaks' AI detector page claims over 99% accuracy and a 0.03% false positive rate. Originality.ai's own accuracy study is headlined "99%+ accuracy" and describes an Academic model with a false positive rate under 1%.

None of those figures were produced by a disinterested party, and none used the same texts, so they cannot be compared with each other.

What the peer-reviewed comparisons found

Weber-Wulff and colleagues (2023), published in the International Journal for Educational Integrity, ran 54 documents through 14 tools including GPTZero and Turnitin. Turnitin scored highest on every accuracy measure. GPTZero did not reach the 70% accuracy that five tools cleared, and half of its positive classifications on human-written or machine-translated documents would have been false accusations, which the authors called unsuitable for academic use. Across all tools, accuracy on machine-paraphrased AI text fell to 26%.

Perkins and colleagues (2024) tested seven detectors, including Turnitin, GPTZero and Copyleaks, on 114 samples. Copyleaks detected 64.8% of unmodified AI text, Turnitin 61%, and GPTZero roughly 26%. After the authors applied evasion techniques, Turnitin's accuracy fell by 42.1 points, the largest drop of any tool, while Copyleaks stayed the most accurate at 58.7%. Copyleaks also flagged half of the small human control set as AI.

Liang and colleagues (2023), published in Patterns, tested seven detectors including GPTZero and Originality.ai on essays by non-native English writers. On average they flagged 61.22% of the human-written TOEFL essays as AI, and 89 of 91 were flagged by at least one tool. That bias has its own thread.

Four detectors, what they claim and what independent testers found

Tool	What it measures	Vendor accuracy claim	Independent finding (source)	Known weak spot
GPTZero	Sentence-level deep learning classifier; human, mixed or AI with a confidence band	False positives no more than 1%; 96.5% on mixed documents	Lowest detection rate of seven tools, about 26% (Perkins 2024); half of positives on human or translated text would be false accusations (Weber-Wulff 2023)	Non-native and translated writing; short passages
Turnitin	Overlapping segments of prose scored 0 to 1, aggregated to a document percentage	Document false positive rate under 1% for documents over 20% AI	Highest accuracy of 14 tools (Weber-Wulff 2023); 61% detection, falling 42.1 points after evasion edits (Perkins 2024)	Paraphrased or lightly edited AI text; needs 300 words of prose
Copyleaks	Linguistic and statistical patterns plus deep learning; percentage of text likely AI	Over 99% accuracy; 0.03% false positive rate	Highest detection at 64.8%, still best after evasion at 58.7%, but flagged 50% of human control samples (Perkins 2024)	False positives on human writing in that test
Originality.ai	Classifier giving "Likely AI" or "Likely Original" with a confidence percentage	99%+ accuracy; Academic model under 1% false positives	One of seven detectors that misclassified most non-native TOEFL essays (Liang 2023); not included in the other two studies	Non-native English writing; results vary by model version

What that means in practice

• Rank order is unstable. Turnitin won one study and dropped to fifth in another once the text was edited. The accuracy thread covers why a single number never tells the story.
• GPTZero's weakness is the costly one. Missing AI text is a leak; flagging human text is an accusation. Weber-Wulff and Liang both found GPTZero on the wrong side of that trade.
• Every study warns against sole reliance. Perkins concluded the tools cannot currently be recommended for deciding whether misconduct occurred; Turnitin's own FAQ says the same of its percentage.