Often, but less consistently than "99%+" suggests. Originality.ai has been tested by more independent teams than almost any other commercial detector, several of them ranked it among the most accurate tools they tried, and it rarely flags everyday human writing. Its own accuracy page puts its current models at "99%+ accuracy" with model-by-model false positive rates between 0.5% and 3%.
The weak spots are just as well documented. In a 2025 University of Chicago study, at a fixed score cut-off it missed 16% of GPT-4.1 text and 37% of Claude Opus 4 text. In the RAID benchmark it flagged 13% of human-written Wikipedia text while its overall false positive rate was held at 5%, and a character-swap attack dropped its accuracy from 85.0% to 9.3%. In Stanford's 2023 data, an earlier version flagged 69 of 91 essays written by non-native English speakers. Below: what each source measured, and which version of the tool it measured.
Originality publishes more numbers than most vendors. The useful question is what each one was measured on.
Originality's accuracy page states that its current models, Originality.ai 5.0 and 5.0 Max, "have 99%+ accuracy." Earlier models are listed with their own false positive rates: under 1% for Lite 1.0.0 and 0.5% for later Lite versions, under 3% and then 1.5% for Turbo 3.0.x, under 1% for Academic, and 2.4% for the Multilingual model. Its AI checker page describes the tool as "the Most Accurate AI Detector in third-party studies."
Most of those figures come without a named corpus, sample size or generator mix, so they cannot be checked. The exception is the 5.0 launch post, which reports specificity (the share of human text correctly called human) on two named public datasets: 99.954% on 25,996 PERSUADE 2.0 student essays, and 98.901% (5.0) or 100% (Max) on the 91 TOEFL essays from the Stanford study discussed below. The post itself notes that its competitor comparisons "come from separate evaluations with different samples and scoring rules; they are not a matched head-to-head test."
Originality is unusually direct here. The AI checker page says "An AI score shouldn't be used alone for academic integrity discipline" and "Use an AI score as one signal, not a final decision." Its founder, Jon Gillham, has written that the company has always recommended against using its tool, or any AI detector, for academic discipline. The tool is built for publishers and content teams, and it requires at least 100 words.
Sources: originality.ai AI checker page, accuracy page (dated October 20, 2026), "Introducing Originality 5.0 and Max" post, and "Do AI detectors work?" post, all read October 2026. Vendor figures, not independent measurements.
Six independent evaluations name Originality with tool-specific results. Read each one with its date, because the tool tested in 2023 is not the tool sold today.
RAID is the largest shared detector benchmark, roughly 6 million texts across 11 generators, 8 domains and 11 adversarial attacks. The authors queried Originality's v1 API on January 24 and 25, 2024. Three findings matter:
Further reading: Is ZeroGPT Accurate? (Lincoln test), on Substack.
| RAID condition (accuracy at a 5% false positive rate) | Originality |
|---|---|
| No attack | 85.0% |
| Paraphrased | 96.7% |
| Synonym swap | 96.5% |
| Misspellings | 78.6% |
| Articles deleted | 71.4% |
| Homoglyphs (look-alike characters) | 9.3% |
Paraphrasing, which breaks many detectors, actually raised Originality's score. Swapping letters for visually identical characters from other alphabets collapsed it. And when every detector was calibrated to a 5% overall false positive rate, the false positives were not spread evenly: "Originality has a 13% FPR on Wikipedia." Encyclopedic prose is close to the register of a school report, which is why that number matters. The authors' summary: Originality "achieved high precision in some constrained scenarios."
Source: Dugan et al., "RAID: A Shared Benchmark for Robust Evaluation of Machine-Generated Text Detectors," ACL 2024 (arXiv:2405.07940). Tables 5 and 6 and Appendix C. Tested version: Originality v1 API, January 2024.
Liang et al. ran seven detectors, Originality.ai first among them, on 91 TOEFL essays by non-native English speakers and on 88 essays by native US eighth-graders, in March 2023. The paper publishes a seven-detector average of 61.3% false positives on the TOEFL essays. The per-detector results are in the authors' public data, and we recomputed Originality's from it: at a 0.5 score cut-off it flagged 69 of 91 TOEFL essays (76%) as AI, against 1 of 88 eighth-grade essays. After the TOEFL essays were polished by GPT-4 for vocabulary, it flagged 11 of 91.
That is a 2023 model. Originality now reports 98.901% specificity on the same 91 essays for its 5.0 model. Because the essays are public, that claim can be checked, but we have not seen an independent team re-run it.
Source: Liang et al., "GPT detectors are biased against non-native English writers," Patterns, 2023, and the authors' data repository (Weixin-Liang/ChatGPT-Detector-Bias on GitHub), OriginalityAI.json files for the TOEFL and Hewlett sets.
On 1,992 human passages across six everyday genres, Originality's false positive rate ranged from 0.27% to 0.11% across score thresholds, second only to Pangram (0.10% to 0%) and ahead of GPTZero (0.71%).
It missed more AI text than either. At fixed score cut-offs between 0.1 and 0.9, Originality missed 11% to 22% of GPT-4.1 passages, 30% to 42% of Claude Opus 4 passages and 13% to 24% of Claude Sonnet 4 passages, while Pangram and GPTZero each missed under 4% of every model's text. The authors' own summary: "OriginalityAI performs worse than both detectors" on missed AI text. When they instead tuned each detector's threshold separately for every genre, Originality's miss rate fell to 5% or less, but that is tuning done after the fact, on the answers, which no ordinary user can do.
On AI text humanized by StealthGPT, Originality missed between 2% and 21%, far fewer than GPTZero (26% to 77%) though more than Pangram (0% to 5%).
Source: Jabarian & Imas, "Artificial Writing and Automated Detection," NBER Working Paper 34223, 2025.
Sources: Walters, doi:10.1515/opis-2022-0158; Liu et al., doi:10.1007/s40979-024-00155-6; Hadra et al., doi:10.1007/s40979-026-00213-1. Quotes from the published abstracts.
A fair audit names the cases where the tool earns its reputation.
Walters and Liu both found high detection rates, Liu including AI text rephrased with another tool. In RAID, Originality was one of the few detectors that paraphrasing made more accurate rather than less, and in the Chicago study it caught far more StealthGPT-humanized text than GPTZero did.
The Chicago study measured false positive rates around a quarter of a percent or lower on reviews, blogs, news and fiction. That is a strong result, and it was measured by a team with no stake in the outcome.
Specificity on named public datasets, an explicit warning against using scores for academic discipline, and an old open benchmark repository on GitHub. You can disagree with the headline numbers, but several of them can at least be located.
Five cases where "99%+" does not carry over. If your writing sits in one of them, treat any score as provisional.
In the Chicago study, at a fixed score threshold, Originality missed 30% to 42% of text written by Claude Opus 4 and 11% to 22% of GPT-4.1 text. A clean score on a document is weaker evidence of human authorship than the headline figure implies, particularly for text from models released after the detector was trained.
RAID's 13% false positive rate on Wikipedia is the clearest single warning. Formal, neutral, information-dense writing is statistically close to what language models produce, and it is also how many students are taught to write reports.
The 2023 Stanford data showed the problem at its worst: 76% of TOEFL essays flagged. Originality says its 5.0 model has fixed that on the same essays. A 2026 journal study still found a borderline accuracy gap between EFL student writing and professional writing. Until an independent team re-tests the current model, treat a flag on non-native writing with particular care.
Hadra et al. found both Originality and Turnitin "performed poorly on Hybrid texts." Most real cases are hybrids: a draft someone wrote and then tidied with a tool, or AI output someone rewrote.
RAID's homoglyph attack took Originality from 85.0% to 9.3%. Vendors patch specific attacks over time, so the exact number may have changed. The general lesson has not: a score can be gamed in ways that have nothing to do with who wrote the text.
If you are asking because your own writing was flagged, this is the table to read. The rate depends heavily on what kind of writing was tested.
| Source | What was tested | False positive rate |
|---|---|---|
| Originality (vendor) | Current and past models, corpus mostly unnamed | 0.5% to 3% by model |
| Originality (vendor) | 5.0 model, 25,996 PERSUADE student essays | about 0.05% |
| University of Chicago, 2025 | 1,992 everyday human passages | 0.27% to 0.11% |
| RAID, January 2024 | Wikipedia, at a 5% overall rate | 13% |
| RAID, January 2024 | Lowest rate achievable on the full set | 0.62% |
| Stanford, March 2023 | 91 TOEFL essays (non-native writers) | 76% |
| Stanford, March 2023 | 88 US eighth-grade essays | about 1% |
The spread runs from a twentieth of a percent to three quarters of the essays, depending on the writer, the genre and the model version. That is why a single false positive rate quoted without its corpus tells you very little about your own document.
A detector score is a reason to look closer, not a finding. This is the sequence that works for people who wrote their own work and were flagged anyway.
Get in writing which detector produced the score, which model or mode it ran, and what score the institution or client treats as actionable. If no threshold is documented, that is worth knowing early.
Two detectors built on different signals disagreeing on the same passage is meaningful and easy to show. Run the identical text through a second tool and keep both results side by side, with timestamps.
Version history is the strongest evidence a writer has, and it is the one thing no detector can manufacture: Google Docs revision history, Word autosaves, notes, outlines, browser history on your sources. A document that visibly grew over many sittings did not arrive from a chatbot.
Most people making these decisions have never read a detector evaluation. Cite the published record rather than arguing about your own score. We keep an appeal-letter template that does this properly.
We sell a detector, so read this section with the scepticism it deserves.
We do not publish a single headline accuracy percentage. One number across every length, genre and register is the move this page has spent its length examining, and it would be no more meaningful coming from us.
What we do publish is one false-positive measurement with the data attached: 1,180 human-written academic abstracts from 2018, before modern chatbots existed, so every AI flag in that set is a false positive by construction. Our detector flagged 5.85% of them, and the per-document results are downloadable so you can check the arithmetic. That figure covers academic abstracts only. It does not tell you how our detector, or Originality.ai's, behaves on student essays, blog posts or fiction, and we do not claim it does.
The detector is English-only, and we have no independent third-party evaluation yet. Originality has been independently tested far more than we have, and the record above is the fairest summary we can give of it.
Sources: our false-positive benchmark (dataset and reproduction script linked on the page) and our accuracy methodology.
No, and Originality says so itself: its AI checker page states that "An AI score shouldn't be used alone for academic integrity discipline." The tool is built for publishers and content teams. Independent tests show it is strong on raw AI text but weaker on hybrid human and AI writing, on formal encyclopedic prose, and, in 2023 data, on essays by non-native English speakers.
It depends on the writing tested. Originality reports 0.5% to 3% depending on the model, mostly without naming a corpus. The University of Chicago measured 0.27% to 0.11% on 1,992 everyday human passages. The RAID benchmark measured 13% on Wikipedia text when the overall rate was set to 5%. Stanford's 2023 data shows an earlier version flagging 76% of TOEFL essays by non-native writers.
It depends on the error you care about. In the University of Chicago NBER study, which tested both on the same text, Originality had fewer false positives (0.27% to 0.11% against 0.71%) and missed far less humanized text, but at fixed score cut-offs it missed much more unedited AI text: 11% to 42% depending on the model, against under 3% for GPTZero. In RAID, GPTZero was more robust to adversarial attacks such as homoglyphs, while Originality was stronger on paraphrased text. Neither wins every test.
One 2026 journal study tested both on the same 192 texts and found Originality's overall accuracy at 0.69 against Turnitin's 0.61, with both performing poorly on hybrid human and AI texts. Walters (2023) found both, along with Copyleaks, had high accuracy on its 126 essays. Turnitin is sold to institutions, not individuals, so most people cannot run the comparison themselves.
An earlier version did, heavily: in Stanford's March 2023 data it flagged 69 of 91 TOEFL essays (76%) as AI. Originality now reports 98.901% specificity on those same 91 essays for its 5.0 model, which would mean about 1 flagged. That figure is the vendor's own and has not been independently re-tested as far as we can find.
Often, yes. In RAID, paraphrasing raised its accuracy from 85.0% to 96.7%, and the University of Chicago study measured it missing 2% to 21% of StealthGPT-humanized text. But RAID also showed a homoglyph attack, swapping letters for look-alike characters, dropping it to 9.3%. Robustness depends on the specific attack.
Originality's Pro plan is listed at $14.95 a month, or $12.95 a month billed yearly, for 2,000 credits a month, where one credit scans 100 words. Pay-as-you-go credits are also sold and expire two years after purchase if unused. Prices were read from originality.ai in October 2026 and may change.
The two detectors most often compared for publishing, side by side on the same evidence.
See the comparison →If credits, pricing or the publisher focus do not fit, the alternatives compared.
See the alternative →The detector that beat Originality on false positives in the Chicago study.
Read the audit →Why human writing gets flagged, and which kinds of writing carry the risk.
Read the guide →1,180 human-written abstracts, our flag rate, and the data to check it.
Check our numbers →Cite the published studies instead of arguing about a percentage.
Use the template →Scan the same passage Originality.ai scored and compare the two readings yourself. 3 checks a day, no account, and the one false-positive benchmark we quote is published with its data.