Ask a detector whether a piece of writing came from a machine and it hands back a percentage. Turnitin, GPTZero and Originality.ai each publish an accuracy figure, and none of those figures are measured the same way, so the useful questions are what each detector actually measures, how often it calls human writing artificial, and what happens to the document you paste in.
TL;DR
- Turnitin publishes a document false positive rate of less than 1% for documents with 20% or more AI writing, and roughly 4% at the sentence level (Turnitin).
- GPTZero claims a 99% accuracy rate and a 1% false positive rate on AI versus human samples, while the same page says “No tool can honestly claim to be 100% accurate” (GPTZero).
- Originality.ai’s current 5.0 models claim 99%+ accuracy, and its published model history records a false positive rate for each version, from 0.5% on the June 2025 Lite model to under 3% on the October 2024 Turbo model (Originality.ai).
- The clearest failure mode is human writing being called AI, not AI text slipping through. In a 2023 study, 89 of 91 TOEFL essays were flagged by at least one of seven detectors (Liang et al.).
- Your text leaves your machine. Originality.ai says scan history may be used to train its models unless you opt out, and that some features send submitted text to OpenAI’s API, where it may sit in abuse-monitoring logs for up to 30 days (Originality.ai privacy policy).
- Turnitin will not score a document under 300 words (Turnitin).
What a detector actually measures
None of these tools tells you who pressed the keys. GPTZero’s technical explainer describes perplexity as “a surprise meter when it comes to AI content detectors”, where “If the text is too predictable, it’s likely to be AI authorship”, and burstiness as “a measure of how much the perplexity varies over the entire document” (GPTZero). The same page notes the company moved to a deep learning architecture in autumn 2023, beyond those two factors, and now runs a sentence-level classifier alongside internet text search and a shield aimed at tools built to defeat detectors. Originality.ai describes a transformer model trained on mixed human and generated text that scores a document and then highlights individual sentences, and is careful about what the number means: “30% Likely AI means the system estimates that roughly 30% of the text is AI-generated, not that it is only 30% confident” (Originality.ai).
Turnitin’s own product description is blunter about the limit of the output: “Its less than 1% false positive rate ensures students are not falsely accused of misconduct” (Turnitin). The tool supplies data. A human decides what it means, and that division is where the failures live.
The numbers each vendor publishes
Turnitin’s figures are the most carefully qualified. In March 2023 its Chief Product Officer defined the problem plainly: “A false positive refers to incorrectly identifying fully human-written text as AI-generated”, alongside a target of “a less than 1% false positive rate” (Turnitin). By May 2023 the company had run 800,000 pre-ChatGPT academic samples through the detector, added an asterisk to any score under 20%, and raised the minimum length from 150 to 300 words, because “Results show that our accuracy increases with a little more text” (Turnitin). The same post breaks down where sentence-level false positives sit: 54% are next to real AI writing, 26% two sentences away, 10% three sentences away, and the remaining 10% nowhere near it. The scale of the problem has also moved. In February 2026 Turnitin reported that “an average of 14.8% of English language submissions to its latest version of the AI detection tool had 80% or more AI-generated writing between October 2025 and February 2026”, against 3.3% in the April to August 2023 window when the original detector launched (Turnitin).
GPTZero publishes several numbers at once. Its October 2026 explainer states that “GPTZero has a 99% accuracy rate and 1% false positive rate when detecting AI versus human samples”, and adds that “No tool can honestly claim to be 100% accurate” (GPTZero). On RAID, described there as 672,000 texts spanning 11 domains, 12 LLMs and 12 adversarial attacks, it “correctly identified 95.7% of AI-written text while misclassifying only 1% of human writing”, with accuracy above 99% on models like GPT-4. It claims 96.5% accuracy on mixed documents, where human and AI writing are combined. The same page concedes that “Detectors can also struggle with unconventional or highly creative writing styles that fall outside the typical training data”.
Originality.ai publishes a model-by-model history. Its current 5.0 and 5.0 Max models, released in October 2026, claim 99%+ accuracy, and earlier entries record false positive rates of 2.8% for the February 2024 Turbo model, under 1% for the July 2024 Lite model, 2.4% for the May 2025 multilingual model and 0.5% for the June 2025 Lite model (Originality.ai). The company is equally candid that its comparison table is not a controlled contest: results from different evaluations “are not a matched head-to-head test” (Originality.ai). Pricing starts free at 10,000 words, with a Pro tier at $12.95 per month billed yearly, or $14.95 month to month, covering 200,000 words a month (Originality.ai pricing).
Where the numbers break
The clearest weakness is not a detector being beaten. It is a detector misreading a person. The 2023 paper “GPT detectors are biased against non-native English writers” ran 91 TOEFL essays through seven detectors: 89 of them, 97.80%, were flagged by at least one detector, and 18 were flagged by all seven. The average false positive rate on that TOEFL set was 61.22%. Rewriting the essays with more native-sounding word choices dropped the false positive rate by 49.45% to 11.77% (Liang et al.). The same paper found that simplifying native-speaker eighth-grade essays pushed false positives the other way, from 5.19% to 56.65%. Our own reporting on how detection is failing inside classrooms is here.
Short text is the second weak point, and each vendor handles it differently. Turnitin refuses to score under 300 words. Originality.ai says “Performance varies by text length, so we evaluate short-text detection separately” (Originality.ai). GPTZero puts it plainly: “Short passages are more difficult to analyze because they offer fewer linguistic patterns for the model to evaluate” (GPTZero).
What happens to the text you paste in
Originality.ai’s privacy policy is specific: “we do not use your personal data or any Student Data to train AI or large language models”, and every user can opt out of having their scan text used to improve the models. The opt-out matters, because “Unless you opt out, we may use your scan history and scan results to help train and improve our models.” Some features route text to OpenAI’s API, and the policy notes “OpenAI may retain submitted text and associated outputs in abuse-monitoring logs for up to 30 days” (Originality.ai privacy policy).
GPTZero’s policy, last updated 2 April 2026, states that “Text and file contents submitted by API subscribers directly by API are not stored by default nor used for product improvement”, with API customers able to opt in to storage so they can view submissions in the app (GPTZero privacy policy).
If the text should not leave the machine at all, a local model exists. Hello-SimpleAI/chatgpt-detector-roberta is published as a transformers text-classification model built on roberta-base and trained on HC3, the corpus of paired human expert and ChatGPT answers collected across open-domain, financial, medical, legal and psychological questions (model card, paper). The card states the model was trained on the whole dataset with no held-out split, for a single epoch, which means any score the authors report from it is optimistic. It is also a narrow base: the only training signal described is human versus ChatGPT writing from early 2023.
If a detector flags your writing
Turnitin itself does not treat a flag as a finding. It says it does not determine misconduct, “rather, we provide data for educators to make an informed decision based on their academic and institutional policies”, and that instructors “will need to apply your professional judgment, knowledge of your students, and the specific context surrounding the assignment” (Turnitin).
Some institutions have gone further than that. Vanderbilt University disabled Turnitin’s detector “for the foreseeable future” after reviewing its reliability, methodology and privacy implications, citing the company’s 1% false positive rate and noting “Turnitin gives no detailed information as to how it determines if a piece of writing is AI-generated or not”. Their conclusion: “we do not believe that AI detection software is an effective tool that should be used” (Vanderbilt University). Their replacement advice is practical rather than technical: talk to the student, compare the submission against their previous work, and check for fabricated sources and factual errors.
What this means
Every vendor number above is measured on that vendor’s own sample, and the vendors themselves say so. Read them as product claims, not as a shared accuracy scoreboard. The part that generalises is the shape of the failure: detectors key on statistical patterns that can track authorship without proving it, so the risk concentrates where a person’s writing sits away from the training distribution - non-native English, or a piece too short to measure well.
That makes the practical decision about a detector less about accuracy and more about what happens next. If the output triggers a grade, a claim or a refund, the score is a starting point for a conversation, and the burden of proof sits with whoever acts on it. If you only want a rough signal on your own writing, a hosted tool will do it, and you can keep the text on your machine by running a local model instead.
The bottom line
Turnitin, GPTZero and Originality.ai all work, in the narrow sense that each is built to separate AI-written text from human-written text and each publishes a headline accuracy figure for that job. None of them can tell you a person did not write it, and each one publishes numbers that only hold inside its own test set. Turnitin is the most transparent about its own limits, down to a sentence-level false positive rate of roughly 4% and a 300-word floor. GPTZero publishes several numbers at once and gives the fullest description of what its benchmark actually contains. Originality.ai publishes the most granular history and the clearest admission that cross-vendor comparison is not a settled thing.
Check the length and the language before you trust a score, keep the text local if it is sensitive, and treat any flag as a question rather than a verdict.