Originality AI is a Canadian-built text checker that scores writing for AI generation and plagiarism, aimed at publishers, agencies and educators — with vendor-reported accuracy figures and terms that forbid using its scores as the sole basis for penalising anyone.
Used this tool? Rate it
Used this tool? Rate it
Originality AI is a web-based text checking service operated by Originality AI Inc., a Canadian company registered at 64 Hurontario St in Collingwood, Ontario. It takes a piece of writing and returns two principal judgements: a score indicating how likely the text is to have been generated by an AI model, and a plagiarism check against existing published material. Around those sit a set of secondary checks — grammar, readability, fact checking, content quality scoring and guideline compliance.
The company was founded by Jon Gillham after he sold a content marketing agency, and the service launched in November 2022 — notably before ChatGPT's public release, when it was built to detect GPT-3 output. That timing is the basis for its claim to have been the first commercially available AI detector.
Everything else in this article depends on one distinction, so it belongs up front. The homepage describes the product as "the Most Accurate AI Detector in third-party studies." The accuracy figures the company publishes — 99%, 99%+, 97.8%, false positive rates from 0.5% to 2.4% — are drawn from datasets its own benchmark tables label "Internal Benchmark." They are vendor-run tests on vendor-selected data, and none of them has been independently audited. Meanwhile the peer-reviewed literature that does exist on this problem reaches a markedly different conclusion about detection tools as a class. Both sides of that are set out below, because a reader deciding whether to rely on this tool needs both.
Three detection models are current, released in September 2025, and the vendor publishes a specific claim for each. Lite 1.0.2 is stated at 99% accuracy on leading flagship models — OpenAI, Gemini, Claude and DeepSeek — with a false positive rate of 0.5%. Turbo 3.0.2 is stated at 99%+ on the same models, with up to 97% accuracy on "humanized" content and a 1.5% false positive rate. Academic 0.0.5 is stated at 99%+ with under 1% false positives, focused on academic material including STEM answers, code and formulas, and up to 92% accuracy against AI humanizer and bypasser tools.
These are the numbers the marketing rests on, and they are worth reading alongside the source the vendor gives for them. The published benchmark table lists Lite 1.0.2 at a 98.91% true positive rate, 0.52% false positive rate and 1.09% false negative rate — against a dataset named, in the vendor's own column heading, "Internal Benchmark: Recently Released Models."
Introduced in June 2026, AI Allowance is a departure from the binary question every detector has asked until now. Rather than "is this AI or not," you set how much AI involvement you are willing to accept — 0%, 5%, 15%, 25% or 40% — and the tool scores against that threshold. The vendor claims 99%+ accuracy at the default 15% setting.
Conceptually this is the most interesting thing in the product, because it acknowledges something the binary framing obscures: most real-world writing in 2026 sits on a spectrum between fully human and fully machine, and an editorial policy usually cares about degree rather than presence. Whether the underlying measurement supports that finer distinction is, again, attested only by internal testing.
The Multilingual 2.0.0 model, released May 2025, covers 30 languages with a stated overall accuracy of 97.8%, false negatives reduced to 1.99% and a false positive rate of 2.4%.
That last figure deserves attention, because it is the vendor's own disclosure that non-English detection is measurably weaker: 2.4% false positives against 0.5% for Lite and under 1% for Academic. Anyone evaluating text written in a language other than English is working with roughly two to five times the false positive rate of the English models, by the company's own numbers.
Plagiarism detection is a co-equal product line rather than an add-on — the founder's stated motivation included building a modern plagiarism checker with scan history, shareable results and team access. Around these two pillars sit the Fact Checker, Grammar Checker, Readability Checker, Content Quality Score and Guideline Checker, plus Bulk Scan for volume work and full site scans that take a URL and crawl it.
The Chrome extension runs scans directly inside Google Docs and can replay the writing process — showing how a document was composed over time rather than judging only the finished text. This is a meaningfully different kind of evidence from a detection score: process evidence rather than statistical inference, and considerably harder to fake. For an educator trying to establish whether a student wrote something, a replay showing organic drafting is more informative than any percentage.
The company has done more than most competitors on transparency: in August 2023 it released an open source benchmark dataset and open source efficacy testing tools. That is a genuine and unusual step in a category where most vendors publish nothing.
It should be weighed against two countervailing facts. The current flagship models' accuracy figures still come from internal benchmarks rather than that open dataset, and the terms of service prohibit reverse engineering, decompiling or attempting to derive the models, weights, prompts or non-public benchmarks — which means outside parties cannot independently verify how the current system reaches its conclusions.
The Enterprise tier is explicitly positioned for agencies and publishers, and this is the use case where the tool's limitations matter least. When an editor is checking whether a freelancer delivered what was commissioned, a detection score is one input into a conversation, not a verdict with consequences. Bulk scan and full site scans make it practical at volume.
AI Allowance fits organisations that have written an actual policy — "AI assistance is acceptable for research and outlining but the prose must be yours" — rather than a blanket ban. Setting a threshold that matches the policy is more honest than pretending a binary detector can enforce a nuanced rule.
The Academic model, Moodle plugin and student data provisions show this segment is served deliberately. But this is also where the tool's own terms impose the sharpest limits, discussed at length below. Used as one signal among several — alongside a writing replay, a draft history, or a conversation with the student — it can contribute. Used as the deciding factor, it violates the vendor's own acceptable use policy.
Anyone commissioning content, evaluating a guest post, or assessing whether product reviews are genuine has a legitimate need here, and the consequences of a false positive are proportionate — you decline a submission rather than penalise a person.
An often-overlooked use: writers who compose in a formal, structured register — which describes a great many non-native English writers, technical writers and academics — can check whether their prose is likely to be flagged, and know in advance if they may need to explain themselves. This is a defensive use of the tool that its critics rarely mention.
Three scans per day of up to 2,000 words are available without an account. Use them on text whose provenance you already know for certain — something you wrote yourself, and something you know was machine-generated — to calibrate your expectations before committing money.
Be aware of the trade-off: text entered into the free detector may be used for training. Opting out requires a paid account. Do not paste confidential or client material into the free tool.
Lite, Turbo and Academic are not interchangeable. Academic is built for STEM answers, code and formulas and reports the lowest false positive rate. Turbo is the more robust option against humanizer tools. If your text is not in English, the Multilingual model applies and carries the higher 2.4% false positive rate.
If you are using AI Allowance, set the percentage to match a policy you have actually written down. Choosing 0% because it feels rigorous, when your real policy tolerates AI-assisted research, produces flags you will then ignore — which trains you to discount the tool entirely.
This is the step that determines whether you use the tool responsibly. A high AI score means the text has statistical properties resembling machine generation. It does not establish how the text was produced. Formal register, simple vocabulary, heavily edited prose and non-native phrasing all push scores upward for reasons unrelated to AI use.
If the result could affect someone, collect other signals: version history, the writing replay, an earlier writing sample for comparison, or a conversation. This is not optional caution on my part — it is what the vendor's terms require of you, as set out below.
One credit equals 100 words. Subscription credits expire at the end of each month and do not roll over; separately purchased pay-as-you-go credits last two years. If your workload is uneven, the pay-as-you-go route avoids paying monthly for capacity you lose.
Enterprise gives 365 days of scan history; Pro does not specify an equivalent retention. If detection results ever inform decisions about people, the ability to revisit what was scanned and when is worth the tier difference on its own.
Section 9 of the terms of service prohibits using output as the sole basis for consequential decisions and prohibits penalising a student, employee, contractor, writer or applicant solely because of a detection or plagiarism score. If you are designing an institutional workflow, that clause should shape it. It is unusual and creditable for a vendor to write this down; it is also binding on you as a user.
Before deploying at scale, run a set of texts you know were written by humans from the same population you will be assessing — same subject matter, same language background, same register. A vendor's benchmark says nothing about your specific corpus, and this is the only way to learn what your own false positive rate looks like.
The vendor's own figures put multilingual false positives at 2.4% against 0.5% for the English Lite model. If you are assessing translated or non-English writing, adjust your confidence accordingly.
Detection reliability improves with length across every published account of how these systems work, including OpenAI's own note that its classifier was unreliable below 1,000 characters. A flag on a paragraph is close to meaningless.
If you have access to process evidence, it is stronger than a score. A replay showing a document being drafted, revised and restructured over hours is affirmative evidence of human authorship in a way no percentage can be.
Free-tier input may be used for training, and only paid accounts can opt out. Client work, unpublished manuscripts and student submissions containing personal data belong in a paid account, if anywhere.
Verified reviewers report the unsubscribe path being difficult to complete. Whatever your view of that, the practical response is to cancel well before a renewal date, confirm in writing, and check the next billing cycle.
Publishers, agencies and content marketplaces are the best fit, because the decisions the tool informs are commercial rather than personal — accepting or rejecting a piece of work — and the tooling around bulk and site-wide scanning matches how they operate.
Editorial teams with a written AI policy get real value from AI Allowance specifically, since a threshold-based tool matches a threshold-based policy in a way a binary detector never could.
Educators and institutions are served deliberately — Academic model, Moodle plugin, student data protections — but should read this article's limitations section in full before building any process around it, and should treat the vendor's own prohibition on sole-basis decisions as a design constraint rather than fine print.
Writers checking their own work, particularly those who write in a formal register or in English as an additional language, have a genuinely defensive use for it.
Who should look elsewhere: anyone seeking proof rather than probability will not find it here or in any competing product, because the underlying problem does not currently admit of proof. Institutions wanting to automate academic misconduct decisions are explicitly outside what the vendor permits. Anyone assessing predominantly non-English writing should weigh the higher stated false positive rate carefully. And anyone who cannot tolerate a small percentage of wrong answers with real consequences attached should not be using automated detection at all.
Originality AI is delivered as a web application with a dashboard, and reaches other environments through a Chrome extension (which runs scans inside Google Docs and provides writing replay), a Moodle plugin for educational institutions, an API for programmatic use, and enterprise integrations. File upload accepts docx, doc and PDF, alongside pasted or typed text, and site scanning takes a URL.
API access is an Enterprise-tier feature rather than being available on Pro. Free tools — the AI checker, fact checker, plagiarism checker, readability and grammar checkers — are exposed as standalone pages on the site.
There are two published subscription tiers, both billed in USD, with annual billing discounted against monthly.
Pro is $12.95/month billed annually ($155.40 per year) or $14.95/month billed monthly, positioned for individuals and small teams. It includes 2,000 credits per month, where one credit equals 100 words — roughly 200,000 words of scanning monthly. Standard support, file upload, full site scans, team management, scan tagging and the Chrome extension are included. Additional team seats are $9.95/month, or $8.62 on annual billing.
Enterprise is $136.58/month billed annually ($1,638.96 per year) or $179/month billed monthly, positioned for agencies and publishers. It includes 15,000 credits per month, a dedicated customer success manager, priority support with issues typically resolved in an hour or less, 365 days of scan history and API access. Additional seats are $24.95/month, or $19.04 annually.
Free access exists without an account: three scans per day of up to 2,000 words. Text submitted this way may be used for training, and only paid accounts can opt out.
Two credit mechanics matter more than the headline price. Subscription credits expire at the end of each month if unused — they do not accumulate. Pay-as-you-go credits, bought as a one-time payment, expire two years from purchase. If your usage is seasonal or project-based, the pay-as-you-go route avoids paying every month for capacity that evaporates.
On refunds, the terms are direct: cancellation stops future renewals but does not automatically refund prior charges, fees are non-refundable once paid unless law or an order form says otherwise, and consumed services — completed scans, plagiarism checks, API calls — cannot be refunded. The terms do preserve any mandatory statutory consumer rights that cannot be waived.
GPTZero is the most direct competitor in the education segment and publishes its own accuracy claims with the same fundamental caveat — vendor-run testing. Comparing the two on your own known-provenance corpus is more informative than comparing their marketing.
Turnitin carries institutional weight through existing LMS relationships and long-standing plagiarism infrastructure, and for universities the integration and procurement path often decides the question before accuracy does. It was one of the commercial systems included in the Weber-Wulff evaluation discussed below.
Copyleaks competes across both AI detection and plagiarism with particular emphasis on multilingual coverage, which may matter if your corpus is largely non-English given the higher stated false positive rates in that setting.
Winston AI and similar newer entrants compete on price and on specific claims about detecting humanizer output. The same evidentiary caution applies to all of them equally.
Doing without automated detection deserves listing as a real option. Process evidence — draft history, version control, oral defence, in-class writing — produces conclusions you can actually defend, and several institutions have moved this way deliberately. It is more work and it scales badly, but it does not carry a false positive rate.
The honest summary: Originality AI is among the more transparent vendors in a category whose core claims are structurally hard to verify. It publishes a model release history, has released open benchmark tooling, and writes explicit limits on use into its own terms. None of that resolves the underlying question of whether any detector is accurate enough for consequential use — and the independent literature suggests the category as a whole is not.
This is the central caveat and it applies without exception. The 99%, 99%+, 97.8% and the false positive rates from 0.5% to 2.4% are all produced by the company's own testing on datasets its benchmark tables label "Internal Benchmark." No independent auditor has verified them. That is not an accusation of bad faith — it is a description of the evidentiary status of the numbers, and it is the normal state of affairs across this entire product category.
The related claim needs stating precisely too. The homepage describes the product as "the Most Accurate AI Detector in third-party studies." The company's own accuracy article contains a section heading promising an overview of third-party studies, but the substantive accuracy data presented on that page comes from internal benchmarks. A reader looking for the third-party studies that would substantiate the homepage claim will not find citations to them there.
The most substantial independent evaluation is Weber-Wulff et al., "Testing of detection tools for AI-generated text," published in the International Journal for Educational Integrity (vol. 19, art. 26, 2023). It tested 12 publicly available detection tools plus two commercial systems, and concluded that available detection tools "are neither accurate nor reliable," with a significant bias toward falsely classifying AI content as human-written. It further found that content obfuscation techniques significantly worsen tool performance, and the authors highlighted serious implications for using such systems in academic settings.
This study did not single out Originality AI, and the models it tested predate the current generation. But it is the closest thing to an independent assessment of the category that exists, and its conclusion is difficult to reconcile with any vendor's 99% claim.
Liang, Yuksekgonul, Mao, Wu and Zou at Stanford published "GPT detectors are biased against non-native English writers" (arXiv 2304.02819, submitted April 2023, finalised July 2023). Their finding: detectors consistently misclassify non-native English writing as AI-generated while classifying native writing accurately. They also showed that simple prompting can both reduce the bias and bypass detection — a result that cuts against detection reliability from both directions.
Originality AI responded publicly, in an article titled "A Response to a Flawed Stanford Study." The response contests the methodology. Notably, it also restates the study's central figures without disputing them: that detectors incorrectly labelled more than half of TOEFL essays as AI-generated, at an average false positive rate of 61.3%.
Readers should draw their own conclusion from that exchange. What is not in dispute is that the risk of misclassifying non-native writing is a documented, published concern that both sides acknowledge exists.
The clearest indication of how hard this problem is comes from the company that builds the models being detected. OpenAI's classifier page carries the note: "As of July 20, 2023, the AI classifier is no longer available due to its low rate of accuracy." Its published performance was that it "correctly identifies 26% of AI-written text (true positives) as 'likely AI-written,' while incorrectly labeling human-written text as AI-written 9% of the time."
That is not a comment on Originality AI's specific models, which claim far better numbers. It is context for how much scepticism the category has earned.
This is the single most important thing in this article for anyone in education or HR, and it comes from the company itself. Section 9 of the terms of service prohibits users from using output "as the sole basis for decisions that could have legal, academic, employment, housing, insurance, credit, disciplinary, reputational, medical, or similarly significant effects on a person," and specifically from acting to "accuse, discipline, penalize, terminate, fail, report, or otherwise materially disadvantage a student, employee, contractor, writer, applicant, or other person solely because of an AI Detection Output or plagiarism score." Section 8 requires that you "use human review and appropriate due process before taking any action that may materially affect a person."
Read plainly: the vendor markets the most accurate detector available and simultaneously forbids you from treating its output as sufficient grounds to fail a student or fire an employee. Both positions are defensible, and holding them together is more honest than most competitors manage. But an institution that adopts this tool to automate misconduct findings is violating the terms it agreed to.
Taking the company's figures at face value, the arithmetic still matters. At the Academic model's stated sub-1% false positive rate, scanning a thousand genuinely human-written assignments could still flag close to ten of them. At the multilingual model's 2.4%, the absolute count is higher again. Percentages that look reassuring in marketing describe a substantial number of individual people when applied across a cohort.
The Trustpilot profile carries a 4.6 rating across 1,294 reviews — a large sample — distributed as 80% five-star, 6% four-star, 1% three-star, 1% two-star and 12% one-star.
Two qualifications. The profile has been claimed by the vendor since September 2023, Trustpilot notes the company actively invites reviews, and the vendor has replied to 89% of negative ones — all of which introduces selection pressure toward the positive. More substantively, reading the recent positive reviews shows they cluster heavily on customer service — repeatedly naming individual support staff — rather than on detection accuracy. The 4.6 is best read as a signal about the company's responsiveness, not as user validation of the detector's correctness.
On the negative side, a recurring specific complaint concerns cancellation. One verified reviewer describes the unsubscribe link as "extremely tiny, well-hidden, and in a font that's almost entirely transparent," followed by around eight retention offers, and states that choosing to submit feedback made it "literally impossible to cancel," leaving the account paused instead.
Subscription credits do not roll over — unused capacity is lost at month end. Cancellation stops future renewals but does not refund prior charges, and consumed scans and API calls cannot be refunded. For uneven workloads the pay-as-you-go option, with two-year expiry, is the better structural fit.
The terms prohibit reverse engineering, decompiling or attempting to derive source code, algorithms, model weights, prompts, system design or non-public benchmarks. This is a standard commercial protection, but combined with internally-generated accuracy figures it means the central claims of the product are not externally checkable by design.
The vendor's own model history shows accuracy on its hardest test set moving from 90.2% to 98.8% across one release, and false positives improving only marginally from 2.9% to 2.8% in the same step. That trajectory reflects a moving target: humanizer and bypasser tools evolve alongside detectors, and a figure published today describes performance against the evasion techniques that existed when the test was run.
The company states 99% or better for its current English models, with false positive rates between 0.5% and under 1%, and 97.8% overall for its 30-language multilingual model at a 2.4% false positive rate. All of these come from the company's own testing on datasets labelled "Internal Benchmark," and none has been independently audited. Independent peer-reviewed work on detection tools as a category — most notably Weber-Wulff et al. 2023 — concluded that available tools are "neither accurate nor reliable." Treat the vendor figures as claims, not measurements.
Yes, and the vendor discloses rates for this directly. More importantly, published research documents that the risk is not evenly distributed: a Stanford study found detectors misclassify non-native English writing as AI-generated far more often than native writing, with an average false positive rate of 61.3% on TOEFL essays. Originality AI disputes that study's methodology while restating its figures. Formal register, simple vocabulary and heavy editing all push scores upward for reasons unrelated to AI use.
Not on its own — and this is the vendor's own rule, not merely advice. The terms of service prohibit using output as the sole basis for decisions with academic, employment or disciplinary consequences, and specifically prohibit penalising a student, employee, contractor, writer or applicant solely because of a detection or plagiarism score. The terms also require human review and appropriate due process before any action that may materially affect a person.
Pro is $12.95/month billed annually or $14.95 monthly, with 2,000 credits per month. Enterprise is $136.58/month billed annually or $179 monthly, with 15,000 credits, API access, 365-day scan history and priority support. One credit equals 100 words. Additional seats cost extra on both tiers. There is a free tier of three scans per day up to 2,000 words, without an account.
Yes — three scans daily of up to 2,000 words, no account needed. The trade-off is that text entered into the free detector may be used for training, and only paid accounts can opt out. Do not use it for confidential or client material.
No, for subscriptions. Subscription credits expire at the end of each month if unused. Pay-as-you-go credits purchased separately expire two years from the purchase date, which suits irregular workloads better.
Yes, via a multilingual model covering 30 languages, with a vendor-stated overall accuracy of 97.8%. Note that the same disclosure puts its false positive rate at 2.4%, against 0.5% for the English Lite model — meaningfully higher, by the vendor's own numbers.
The vendor claims Turbo 3.0.2 identifies humanized content with up to 97% accuracy and Academic 0.0.5 is robust against humanizer and bypasser tools with up to 92%. These are internal figures. Detection and evasion co-evolve, so any such number describes performance against the bypass techniques available when the test was run.
The Trustpilot distribution is 4.6 across 1,294 reviews with 80% five-star and 12% one-star. Positive reviews cluster on customer service rather than detection accuracy. The recurring negative theme is cancellation difficulty — one verified reviewer describes a near-invisible unsubscribe link, roughly eight retention offers, and being unable to complete cancellation after choosing to submit feedback.
No, and neither can any competitor. A detection score is a statistical likelihood, not evidence of how a text was produced. The most probative thing the product offers is not the score but the writing replay in its Chrome extension, which records how a document was actually composed. Process evidence of that kind is far stronger than any percentage — and it is the direction most defensible academic integrity practice has been moving.
