Originality.ai

By Minel Gunesoglu, founder. For this page I opened Originality.ai's pricing page in both billing views and its Terms and Conditions, the two independent studies that measure this product by name, the benchmark paper its own marketing points at, and its Product Hunt and Capterra review records plus two Hacker News comments cited by item ID, all read between 3 and 9 August 2026. No document of mine was put through Originality.ai and no detection run of any kind sits behind anything below. Re-checked monthly.

The strongest claim Originality.ai makes about itself is supported by the document it points at, and the same document is where the sharpest published limit on the product sits. The vendor's post of 28 October 2025 says it was the most accurate detector on the base dataset of a benchmark called RAID, at 85%. RAID (Dugan and colleagues, ACL 2024) evaluates eight open-source and four closed-source detectors, names this one in its own line listing the commercial four, and carries the 85. The same paper also names Originality.ai among five detectors that "sometimes deteriorate from perfect accuracy to complete failure" once the text generator or the decoding strategy changes. Both halves come from one source, and both are below in that order.

Originality.ai is an AI-text detector sold together with a plagiarism checker, a fact checker and a readability scorer, on a single credit meter where one credit covers 100 words. The vendor's own pricing page describes its cheaper plan as being "For individuals and small teams." and its expensive one as "Ideal for agencies & publishers.". That is the audience this product is built and priced for, and it is not the same audience as the person whose essay came back with a percentage attached.

Two things decide what the numbers on this page are worth. Nothing here was produced by scanning a document: this desk runs no detection tests of any kind, so every accuracy figure below belongs to a named study or to the vendor, carries the conditions it was measured under, and links to the document it came from. And this site sells no detector and no writing tool, and does not cite, link or list products from the category that markets itself against detection. That second sentence is worth stating plainly on a page like this one, because it is the part a reader cannot check for themselves.

Originality.ai Has No Free Tier: What the Paid Plans Cost and How Far a Credit Goes

Originality.ai's pricing page, read in a browser on 9 August 2026 in both of its billing views, carries two plans and no third tile of any kind. There is no free plan, no trial tile and no talk-to-sales tier. In a category where free scans are ordinary, paying is the only way in, and that belongs at the top of a buyer's calculation rather than in a footnote.

The unit underneath both rows is the credit, and the page defines it in five words on each plan: "1 credit equals 100 words". It sets an expiry in five more, also on both plans and in both views: "1 month expiry - renews monthly". Credits do not accumulate. A month you do not use is gone, and an annual subscriber's allowance still arrives and expires monthly rather than pooling into a year.

Apply the vendor's own published rate and the meter becomes concrete. A 2,000-word article draws 20 credits. Pro's 2,000 credits a month cover 200,000 words, or a hundred such articles, before anything is checked twice. Enterprise's 15,000 cover 1.5 million words a month. None of those three word counts appears on originality.ai/pricing. Each is one multiplication away from the rate that does, and each is a monthly ceiling rather than a balance, because the allowance resets instead of rolling forward.

Two things a buyer looks for are worth stating from the same capture. The only price reduction the page offers is the billing toggle, labelled "Annual (Save up to 23%)", which is what produces the $12.95 and $136.58 rows and the "You are saving $24" and "You are saving $509" lines beside them; no coupon field, promotional tier or seasonal price appeared in either view on 9 August 2026. And on running the meter dry mid-month, the contract grants permission rather than describing a mechanism: Section 7 says "Overage or top-up charges may apply if your usage exceeds plan limits." No price for that, no threshold at which it starts and no name for the product it buys appears in the Terms or on the pricing page, so this page states the permission and stops.

One mechanic that circulates widely is deliberately not stated here. Our own earlier note recorded that an AI-only scan costs one credit while a combined AI and plagiarism scan costs two. That mechanic is not on the pricing page in either billing view, and the vendor help article that would document it returns a 404. It may well be true. It is not sourceable today, so this page asserts no per-scan multiplier, and the note was withdrawn on 9 August 2026 rather than quietly kept.

What separates the two plans beyond the credit count: Enterprise carries API access, 365 days of scan history, a dedicated customer success manager and priority support, against Pro's 30 days of history and standard support. Both carry team management, file upload, the Chrome extension and full site scans, which the page sells with the line "Simply enter the URL of any website to scan it.", a feature worth pricing carefully on a meter denominated in hundred-word units, and one a reviewer's complaint further down attaches to directly. For an institution, that Enterprise row is the whole published record: $179 a month, 15,000 credits, and no education-specific pricing, no security-certification list, no learning-management integration inventory and no API rate card appear anywhere in the material captured for this page. Section 1 of the Terms names a Moodle plugin among the services those Terms cover, which is a contract's list of what it governs rather than evidence of what integrates with what.

One structural note for anyone weighing this against Turnitin, which is the comparison readers arrive with: Turnitin reaches most students through an institution's own licence inside a learning-management system, while Originality.ai is bought directly and metered by credit. Before the accuracy question is worth asking, it is worth settling which of those two you already have access to, because they are not the same kind of purchase.

Originality.ai Free Trial: What the Pricing Page Showed on 9 August 2026

Originality.ai's pricing page, read that day in both billing views, showed the two paid plans above and nothing else: no free tier, no trial, no starter credits named on the page. The Terms and Conditions, last amended 17 June 2026, treat offers of that kind as discretionary rather than fixed: "We may offer trials, promotional pricing, beta access, limited free scans, free word limits, or similar offers. We may change, suspend, or discontinue promotional offers at any time."

A second dated record points a different way. A Capterra reviewer's length-of-use field reads "I used a free trial" (Manish G., 13 July 2024). Both facts stand as recorded, two years apart, and this page joins them with nothing. It does not say a trial was withdrawn, and it does not say the reviewer was mistaken. Capterra's length-of-use field is a self-selected dropdown, the pricing capture is a page read in August 2026, and neither document supports a claim about what changed in between. What a buyer can act on is the August 2026 page: to try this product, you pay first.

Originality.ai's Refund Policy: Case-by-Case Discretion, and What Happens to Unused Credits

Originality.ai's Terms take a firm default on money already spent. Section 7 states that "Credits, scan allowances, word allowances, API usage, or similar usage units may expire, be subject to plan limits, or be non-transferable unless we state otherwise in writing", which is the contractual side of the pricing page's monthly expiry. On refunds the same section reads: "fees are non-refundable once paid, and completed scans, plagiarism checks, AI-detection checks, citation checks, fact checks, proofreading outputs, API calls, credit usage, and other consumed Services cannot be cancelled or refunded". The pricing page's own promise is "Cancel anytime", which governs renewal rather than recovery of what has already been billed.

Both clauses open with the same carve-out (unless law, an order form or a separate refund policy says otherwise) and that separate policy exists. Read on 9 August 2026 and last amended 16 July 2025, it makes refunds discretionary rather than scheduled: they "are reviewed on a case-by-case basis, and a partial or full refund may be granted at our discretion" for service disruptions, billing errors, or "Any other reason we see fit". There is no stated window and no automatic entitlement. Cancelling keeps access to the end of the period already paid for, and a refund that has not arrived within 30 days of being applied is the point at which the policy tells you to make contact.

One sentence in it decides whether a refund is worth asking for, and it is the clause a buyer sizing an allowance should read twice: "If a refund is processed for any of the above-noted pricing plans, all remaining credits in the associated account will automatically expire and be invalid." Taking money back means giving the balance up. That is an ordinary way to write a refund term and not unique to this vendor, but it is not on the pricing page, and it changes the arithmetic on an unused annual allowance.

The refund policy also names a plan the pricing page does not. It lists three pricing options ("Pay as you go (one-time payment)" alongside Pro and Enterprise) and gives that first one a different clock: those credits "expire 2 years after the purchase date (if unused)", against the monthly expiry on both subscriptions. The pricing page read the same day showed two plans and no third option of any kind. Both pages are live, the policy is a year older than the pricing page, and this page joins them with nothing: it does not say the plan was withdrawn and it does not say either page is wrong. Anyone with a one-off job rather than a monthly flow should ask which of the two is current before assuming the two-year credit exists to buy.

Which Accuracy Figure Originality.ai's Capability Cell Converts, and Why the Composite Is 52.6

Originality.ai's card reads 52.6 out of 100. Every source behind it was read on 8 or 9 August 2026, and the card carries the older of those dates, so nothing in it is older than 8 August. Anything under 50 is graded low, which ranks the figure and does nothing else. It is not a finding about the company, and it is not a summary of the sections underneath. The card's own shape is the more useful thing to know, and it is not the shape a low figure usually implies. Every dimension carries a reading, and only one of the four rests on a deep source. Value converts a published price list. Capability converts a ten-item arm, usability converts twenty-two reviews, and commercial terms converts two positions in documents the vendor wrote about itself. A card that is nearly full of slight readings is a different object from a half-empty card of solid ones, and no single word above the number tells you which one you are looking at.

Four readings, multiplied, come to 7,638,432.5, and the fourth root of that product is 52.57, which the card publishes as 52.6 and 2.6 on the five-point scale. Averaged flat rather than multiplied, the same four would give 54.3. The gap is a point and seven tenths, and on this card that is the informative part: multiplying punishes one weak cell sitting among strong ones far harder than it punishes four cells that are all somewhere in the middle. Nothing here is high enough for the operation to have much to take away.

The capability cell is the one that needs explaining, because two independent studies measure this product's accuracy and the cell converts only one of them. It converts the 40.0: accuracy on ChatGPT-written submissions after they were paraphrased through QuillBot, from Orenstrakh and colleagues, peer-reviewed at IEEE COMPSAC. The same study puts Originality.ai at 100% on those ten submissions before the paraphrasing, and n=10 is a thin arm, stated as thin wherever it appears.

That study was chosen for a reason that has nothing to do with which number is higher. It is the only one on this page that measures both detectors carrying a score card on this site, on one corpus, by one method, the same 114 pre-ChatGPT human submissions and the same forty ChatGPT ones. Until 9 August 2026 this cell converted a different study, in Arabic, measuring a different kind of degradation, while the sibling card converted this one. Both composites then sat on a category page that orders by score, and a reader had no way to see that the two capability figures were never measured against each other. Our own records said the pair was not a ranking; the layout invited it to be read as one. Moving both cells onto the shared study is what makes the comparison legitimate, and the number moving up is a consequence of that, not the reason for it.

The Arabic study has not left the page. It is reported in full in the next section (92% on clean articles, 12% after light polishing, with Originality.ai named the best-performing commercial model on the clean set) and it remains the sharper finding about what happens when text is edited. What changed is which of two measurements a single cell is allowed to stand for, not what this desk knows.

Usability converts Product Hunt's 3.0 across 22 reviews. Capterra's 5.0 across 7 reviews does not feed it: seven sits below the minimum sample this engine will convert, so it would have been dropped whichever way it pointed. Both numbers appear in prose below, because the disagreement between them is more useful than either alone. G2, Trustpilot and Gartner Peer Insights refused every retrieval method attempted, in separate sessions and by separate tools, so this page prints no rating from any of the three and borrows none from a page claiming to have read them.

Commercial terms converts two positions, and at 34.5 it is the lowest reading in that dimension on any card this desk has published as of 12 August 2026. The first position is Section 21's indemnity, set out at length further down this page: a non-consumer customer indemnifies the vendor over decisions made about a person on the strength of the vendor's own output, while Section 20 excludes the vendor's liability for those same academic, employment and disciplinary consequences. That is the shape the rubric scores at the bottom of its scale, and the rubric describes the clause rather than grading the company. The second position is the refund posture, and it pulls the other way: the refund policy publishes criteria and reviews requests case by case, which is a stated position rather than silence, and it is why this cell reads 34.5 rather than the 20.0 it read until 12 August 2026, when it was one clause on its own. The other two positions do not apply (a detector's output is a verdict about somebody else's text rather than a work to own, and there is no free tier to have free-tier terms about) so both leave the average rather than scoring zero in it. One thing is worth stating once and without grievance: Section 9 of Originality.ai's Terms requires the company's prior written consent before "public comparative benchmarking, performance testing, vulnerability testing, or publication of benchmark results". That is the contract's text, dated 17 June 2026, and it is part of why the accuracy evidence on this page belongs to published researchers rather than to us. Our scoring method is published separately, including the rule that decides when a dimension is left empty instead of estimated, and the rest of the AI-detectors category prints each card's basis beside its figure.

Is Originality.ai Accurate? What the Independent Studies Measured

Three separate measurements of Originality.ai exist in public, taken on different corpora, in different languages, under different conditions. Two are independent academic studies that name the product; the third is the benchmark paper the vendor's own marketing points at. None was produced here, each is named and dated below, and none of them can tell an individual reader what happened to their own document.

Originality.ai on Clean Arabic Text: 92% Before Polishing, 12% After

Almohaimeed, Almohaimeed, Jari, Alobaid and Alotaibi built two Arabic-language datasets and ran fourteen language models and commercial detectors across the first, then carried the best eight forward to the second (arXiv 2511.16690, submitted 16 November 2025, revised 2 December 2025). The first dataset is 800 articles, half AI-generated and half human-authored. The second, Ar-APT, takes 400 human-written Arabic articles and has ten language models polish them across four settings, for 16,400 samples in total, and it is those eight that were measured against it. The study's question is not whether a detector spots AI writing; it is whether a detector starts calling human writing artificial once somebody has run it through a light edit.

On the clean 800-article dataset the study records Originality.ai as "the best performing commercial model", at 92%, and on articles slightly polished by Mistral or Gemma-3 the same product reads 12%. Both figures belong to one paper, and on this page neither travels without the other.

Two pieces of context travel with that pair every time it appears. The scope is Arabic-language articles, polished by Mistral, Gemma-3 and LLaMA-3; it is not an English-language result, not a general claim about the product, and nothing here implies otherwise. And the finding is about the category before it is about one vendor: in the study's own summary, "The results reveal that all AI detectors incorrectly attribute a significant number of articles to AI", with the best-performing language model tested, Claude-4 Sonnet, reaching 83.51% and falling to 57.63% on articles polished by LLaMA-3. Nobody in that study solved the problem.

Two limits apply to the citation itself. This is a preprint, submitted to the Journal of Big Data, and peer review is not confirmed; that status travels with the number wherever it goes. And the abstract states no false-positive rate of any kind. Figures of that shape circulate for this product, none of them comes from here, and only the abstract was read, so this page publishes what the abstract says and stops there.

Originality.ai's RAID Benchmark Claim: 85% on the Base Dataset, 14% Under One Condition

Originality.ai's claim to have led this benchmark is supported by the paper it points at, and that is worth saying plainly before anything else. Writing on 28 October 2025, Jonathan Gillham reported: "Most Accurate AI Detector on Base Dataset: Originality.ai was the most accurate AI detector on the base dataset with 85% accuracy vs the closest competitor at only 80%." The paper carries the 85, names the product in its own list of the commercial systems it evaluated ("Commercial: GPTZero, Originality, Winston, ZeroGPT") and matches the twelve-detector count the post describes. The runner-up figure was not separately checked against the paper here and stands as the vendor's own reading.

The same paper is also where the product's sharpest published limit sits, and the conditions are the whole of it. Table 5 gives Originality.ai 100% on ChatGPT output generated with sampling, and 14% on output from MPT under greedy decoding with a repetition penalty applied. That is a best-case and worst-case pair from one table under stated settings; it is not an accuracy range, and it describes those generators and those decoding choices rather than the product in general. The vendor's own post names what did not go well, in the same breath as the favourable numbers: Originality.ai "placed 1st in 9 of the 11 tests, 2nd in 1 test, and performed poorly on 2 rarely used bypassing techniques which we will be addressing shortly". A page quoting the first half of that sentence and dropping the second would be selecting from a source that had already disclosed both.

Two further readings from the same paper matter more to a reader than the headline does. The first is a floor rather than a collapse: the researchers could not push Originality.ai below a 0.62% false-positive rate in their own sweep (the lowest of the three floors that sentence names, against 0.88% for FastDetectGPT and 16.9% for ZeroGPT) and "plateauing at 0.62%" describes the limit of what the study could produce rather than an error rate in ordinary use. The second comes from the vendor rather than the paper, and the distinction matters enough to mark: Originality.ai's own post about the study says "The study used a 5% False Positive Threshold for all tests." That is the vendor's reading of the benchmark's methodology, not a sentence found in the paper captured for this page, and it is repeated here on that footing. If it holds, then at the setting described, five human documents in every hundred were expected to be flagged: which is the most useful line on this page for anyone holding a percentage about their own writing, and the reason it is worth being precise about who said it.

The paper's own conclusion is about the category rather than this vendor: detectors in its sample do not generalise across different models or generation settings inside the same domain, and its authors record that changing the generator, switching decoding strategies or applying a repetition penalty "was enough to introduce a 95+% error rate". RAID measured 2024-era generators across more than six million generations, and its coverage ends there.

Originality.ai's False-Positive Rate: What a Second, English-Language Study Found

The design is what makes this study usable here: its human corpus predates ChatGPT, which removes the argument about provenance entirely, a flag raised against that text is an error, not a judgement call. Orenstrakh, Karnalim, Suárez and Liut assembled 124 such computing-education submissions, added 40 written by ChatGPT, and put eight publicly available detectors over the pair (arXiv 2307.07411, submitted 10 July 2023, peer-reviewed at IEEE COMPSAC on 2 July 2024, DOI 10.1109/compsac61105.2024.00027). It is also the nearest public equivalent of what a person does privately when they feed an old essay of their own to a detector to see what comes back.

Across 114 of those human submissions, Originality.ai returned seven false positives, sixth of the eight tools measured, on a table whose readings run from zero to fifty-two. Sixth is also where the study's two overall-accuracy measures place it. Its threshold measure records 93.86% against the human corpus and 70.00% against the ChatGPT one; its weighted measure records 86.48% and 66.77%. Four tables later, ten of the ChatGPT submissions were passed through the paraphrasing tool QuillBot and every detector was measured again: Originality.ai fell from 100% to 40% by the threshold measure, and from 90.24% to 40% weighted. Ten is a small enough sample that the fall is better read as a direction than as a rate.

Every figure in the paragraph above comes from a 2023 corpus of computing-education submissions written in English, 124 human papers and 40 generated ones, which cannot speak to any model released since. What it adds is not a verdict but a repetition: the same shape (strong on clean text, sharply worse once that text has been lightly rewritten), turns up in two independent studies, in two languages, under two different kinds of editing, and one of the two cleared peer review. A reader inclined to treat the 12% as one hostile preprint has an answer, and it is a fair one. This study neither confirms nor contradicts the RAID findings above; it is a different corpus asking a different question.

Can Originality.ai Be Wrong? What Its Terms and Its Reviewers Say

Originality.ai answers this question itself, in the contract every customer accepts. Section 4.1 of its Terms and Conditions, last amended 17 June 2026, reads: "AI Detection Outputs are probabilistic estimates. They may produce false positives or false negatives and may be affected by language, genre, paraphrasing, translation, editing, short text length, model changes, mixed authorship, or other factors. You must not treat any AI Detection Output as conclusive proof that a person used or did not use AI." The paragraph above it is shorter: "The Services are decision-support tools. They are not a substitute for professional judgment, editorial review, legal advice, academic due process, instructor discretion, or human evaluation."

Read that clause's own list of what affects a score (paraphrasing, translation, editing), beside the studies in the section above, which measure this product under exactly those conditions. The contract and the research describe the same behaviour. Both documents are public and dated, and between them they establish that a percentage from this tool is an estimate the company selling it declines to call proof.

Originality.ai's Terms: Who Is Liable for a False Positive?

The indemnity in Section 21 is scoped by its own opening words, and that scope is the part most summaries drop: "If you are a business, organization, School Customer, agency, publisher, institution, or other non-consumer user, you will defend, indemnify, and hold harmless Originality.AI and its affiliates, directors, officers, employees, contractors, suppliers, licensors, service providers, agents, successors, and assigns from and against claims, damages, liabilities, losses, costs, and expenses, including reasonable legal fees, arising out of or relating to:" and among the triggers listed beneath it, "your use of Outputs to make academic, employment, disciplinary, legal, commercial, or other decisions about a person".

That obligation runs to non-consumer users. An individual consumer buying a Pro plan is not inside it. A university, an employer or an agency is, and so is the institution that scanned a student's work. If you are the person a score was produced about, this clause is not yours to carry; it belongs to whoever ran the scan.

Section 20 handles the other direction. Originality.ai excludes liability for, among other things, "ACADEMIC, EMPLOYMENT, DISCIPLINARY, REPUTATIONAL, OR COMMERCIAL CONSEQUENCES; OR CLAIMS ARISING FROM RELIANCE ON OUTPUTS, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGES", and caps total liability at "THE GREATER OF: (A) THE AMOUNT YOU PAID TO ORIGINALITY.AI FOR THE SERVICE GIVING RISE TO THE CLAIM IN THE THREE MONTHS BEFORE THE EVENT GIVING RISE TO LIABILITY; OR (B) CAD $100." The same five words (academic, employment, disciplinary, reputational, commercial), appear on both sides of that arrangement: as consequences the vendor is not liable for, and as decisions a non-consumer customer indemnifies the vendor over.

An allocation of that kind is ordinary in enterprise software contracts, and this page draws no conclusion from it about the company, its intentions, or the standing of anything it publishes. What is worth a buyer's attention is where it sits: inside a product sold for making exactly those decisions. Section 9 makes the same point from the user's side, listing among prohibited uses "use 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 "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 10 puts it in one line for education customers: scores and reports "should be treated as indicators requiring review, not as definitive proof of misconduct". The contract is governed by Ontario law, with exclusive jurisdiction in Ontario courts. Those are quotations with a date on them, and they are the most useful sentences in this entire record for anybody being shown a number.

Originality.ai Complaints: What Product Hunt's Tag Counts Show, and How Capterra Disagrees

Originality.ai's two readable review records disagree by two full stars and agree on the direction of the complaint, which is the more interesting half. On Product Hunt the product holds 3.0 out of 5 across 22 reviews, read 9 August 2026. The strongest signal there is not any single review but the platform's own tag counts across all 22: eleven reviewers tagged inaccurate AI detection and seven tagged false positives, on a page where six tagged AI detection as a strength, six named the interface and six the plagiarism detection. Four tagged inaccurate plagiarism detection; one, support. The sample is composed of five reviews in the platform's founder tab and seventeen in the other, and Product Hunt renders relative ages only, so no review there can be given a precise date.

The individual reports run both ways. A reviewer with a single review to their name described a procedure rather than an impression: "I scanned tree documents: one entirely produced with AI, another completely original" [sic], plus a passage written in the moment about their day, and reported that "All showed a horribly 99% AI generated content" [sic]. Another single-review account reported the AI feature returning human-written text as artificial, the plagiarism checker as inaccurate, the readability feature splitting sentences at every full stop, a whole-site scan that consumed credits with no cost estimate and no way to stop it, and no answer from support. Against those, a reviewer in the founder tab wrote of the detector: "It picked up the AI content every time.": posted about three years before this page was read. A review aggregate supports no accuracy rate in either direction, and none of these is a measurement.

Capterra shows the opposite headline: 5.0 out of 5 from exactly seven reviews, every one of them five stars, with ease of use at 4.9 and customer service at 5.0. Two arrived on 9 May 2024 and five between 13 and 17 July 2024, three of those on one day; nothing has been added in the twenty-four months since, which makes it a two-year-old sample rather than a current one. Seven is below the sample floor this engine converts, so none of it feeds the score.

Look past the ratings and the two platforms point the same way. Capterra's only critical line comes from a reviewer who still awarded five stars: "It may be too sensitive when it comes to detecting AI writing content" (Sade P., 13 July 2024). Over-flagging is precisely what Product Hunt's eleven tags describe. Two records, two stars apart, from different years, both far too small to settle anything, disagreeing about the rating and agreeing about the fault line.

The most useful review in either record is positive, and it is not about accuracy at all. Dermot L., an instructor in higher education with more than a year on the product, wrote on 13 July 2024: "I use it in tandem with other tools and my experience to make a final assessment." That is a customer independently describing the posture the vendor's own Terms require, and it is the shape of a workflow that survives a wrong answer. His stated gap is worth carrying too: the product gives him a link he can show a student, but "It doesn't provide a narrative report that I can send to a student."

Two Hacker News comments on one thread supply the only individual-writer material distinct from those platform aggregates, and both rendered as five months old when the thread was captured on 9 August 2026 (Hacker News shows relative ages only, so the item IDs are the citable part. On item 47025122, greenfrogs wrote: "hey! im not op but ive used originality.ai before and it saved my ass. its super sensitive, but also super accurate". Four lines beneath it sits a direct reply from cubefox: "I tested it, I think it's super inaccurate." On item 47026653, the same account reported: "I just tested originality.ai and it claimed 100% probability that the editors note on the Ars retraction"), the comment links a February 2026 Ars Technica editor's note at that point: "was itself AI generated. For the Gemini article on Benji Edwards it was 'only' 56%." Those three quotations come from two accounts rather than three, which is worth knowing before anyone counts them as independent corroboration. One person's report of one run on one document is a user account rather than a measurement: no method, no sample, no verification, and it sits at the opposite end of the evidence scale from the studies above.

One practical note that costs nothing to know. Product Hunt carries two listings for this brand. The one anybody would guess from the name shows two followers and no reviews at all; the record described above lives at the longer slug linked from this page. If you have gone looking for Originality.ai's community record and concluded there isn't one, that may be why.

Alternatives to Originality.ai: Copyleaks Is the Name on This Page

One other AI detector on this site carries a score card, and as of 9 August 2026 that is Copyleaks, built from public sources under the same rules as the card above.

Both cards should be read as separate measurements rather than as a ranking. Their capability cells convert figures from different studies, taken on different corpora in different languages, and nothing has been produced here that would make them comparable. Comparing them properly would take a page built to do that job, and no such page exists here yet.

One difference is worth naming for a reader who has just read that Originality.ai has no free tier, and it is a fact about entitlement rather than a judgement about quality: that card's own detector page, read on 8 August 2026, offers a scan "up to 25,000 characters without logging in". Whether the tool doing the scanning is any good is a separate question, answered there by its own evidence and not by this sentence.

Searches for alternatives to this product return, in quantity, a class of vendor that sells against detection. None of them is named, cited, linked or listed here, as a source or as an alternative. Several detectors are named elsewhere on this page inside quotations, because a paper's own list of the systems it evaluated names them; none of those mentions is a recommendation, and none carries a number from anywhere but the paper itself.

Is Originality.ai Worth It? The Buyer's Answer and the Flagged Writer's Answer

For a buyer sizing real volume, Originality.ai's 52.6 out of 100 is not the number to decide on, because three of its four cells are slight readings rather than deep ones. The decision sits in three facts that are firm. You cannot try this product without paying: there is no free tier and no trial on the pricing page as read on 9 August 2026. The meter is 100 words to a credit, it renews monthly and unused credits expire, so a team with uneven monthly volume pays for capacity it cannot bank. And the accuracy being bought is well documented on clean text and documented as much weaker on text that has been lightly edited, in two independent studies and in the vendor's own contract. If the work is scanning large volumes of other people's submitted copy inside a process where a flag triggers a human look (the use the instructor quoted above describes), the price is ordinary for this category and the product is built for that. If what you want is one clean percentage you can act on without reading anything else, no detector in the studies on this page supplies that, and this one should not be bought in the belief that it does.

For a writer holding a percentage about their own work, four things on this page are worth more than the composite. Originality.ai's own post about the benchmark that produced its best number says that study ran at a 5% false-positive threshold throughout (the vendor's account of the method, not the paper's own words), which on its face means five human documents in every hundred were expected to be flagged at that setting. Section 4.1 of the vendor's contract says the output is a probabilistic estimate that must not be treated as conclusive proof that a person used or did not use AI. Section 9 of the same contract prohibits its own customers from using a score as the sole basis for a significant decision about a person, or from penalising anyone solely because of one. Section 21's indemnity does not bind you as an individual: it binds the institution or business that ran the scan. What tends to help in that conversation is not an argument about the writing but a record set beside it: the authorship trail your own editor already keeps in its draft and version history, and the full report behind the percentage rather than the percentage by itself. Asking which document produced the figure, and asking to see it, is a request for the evidence rather than a dispute with it.

For an institution, the published record is short: $179 a month buys 15,000 credits, API access and a year of scan history, with no education pricing, certification list or integration inventory available publicly. The clause to put in front of whoever signs is Section 21, in full, beside Section 20, and the reason is not that either is unusual, but that this product is sold for making the exact decisions those clauses allocate.

---