Guide

Turnitin AI Detection in 2026: How It Works, How Accurate It Is, What Students Should Know

A plain-English 2026 guide to Turnitin AI detection: how the AI writing indicator scores a document, what the 20% threshold and the asterisk actually mean, which models Turnitin says it can detect, what Turnitin's own accuracy claims cover, what peer-reviewed testing found instead, which universities have switched it off, and what to do if you are flagged.

August 30, 2026 · The AI Rankings

Quick answer: Turnitin AI detection is an “AI writing indicator” that appears inside the Turnitin Similarity Report and estimates what percentage of a document’s long-form prose was likely generated or modified by an AI writing tool. Turnitin’s published claim is a false-positive rate of under 1% on documents where more than 20% of the text is flagged, validated before each model release against 700,000 academic papers written before ChatGPT launched (Turnitin FAQs). The independent evidence in 2026 points the opposite way from the one most people expect: a peer-reviewed study published in June 2026 in the International Journal for Educational Integrity found Turnitin produced zero false positives on fully human papers but scored 0% on 100% of the fully AI-generated papers it was given (Van Vlasselaer et al.). Three things follow. Turnitin only shows the indicator to instructors and administrators, so students cannot pre-check their own work. Scores between 1% and 19% are hidden behind an asterisk because Turnitin considers them too unreliable to print. And Turnitin itself states the percentage “should not be used as the sole basis for action” — a position a New York court endorsed on 28 January 2026 when it annulled an academic integrity finding built on a 100% Turnitin AI score.

This guide covers Turnitin specifically: the product, the score, the evidence and the fallout. For the underlying detection mechanism across all tools, see how AI detectors work. For ranked alternatives with independent accuracy scores, see best AI detectors. For the tools on the other side of the desk, see best AI for students.

What Turnitin AI detection actually is

Turnitin AI detection is a feature of the Turnitin Similarity Report, not a separate product a student or writer can buy.

Four facts about the product shape everything else on this page.

It is licensed, not universal. AI writing detection is only available to institutions that license Turnitin Originality, or to iThenticate 2.0 customers who buy the AI writing add-on (Turnitin FAQs). An institution running plain Turnitin Feedback Studio or Turnitin Similarity may have no AI indicator at all.

Administrators can switch it off, and many have. Turnitin states that admins can disable the feature from their settings page, that submissions are not processed for AI writing when it is disabled, and that scores cannot be generated retroactively — a flagged document has to be resubmitted.

Students never see the score. The indicator and report “are not visible to students”. An instructor can download the AI report as a PDF and share it, but there is no student-facing view and no consumer version of the tool. Anything marketed online as a “Turnitin AI checker” for students is not Turnitin.

It is separate from the similarity score. Turnitin is explicit that the Similarity score and the AI writing percentage “are completely independent and do not influence each other”. A 4% similarity score and a 92% AI score are two different measurements of two different things, and conflating them is one of the most common errors in student-facing advice.

How Turnitin’s detector works

Turnitin splits a submission into overlapping segments, scores each segment between 0 and 1 for the probability it is AI-generated, pools those scores down to individual sentences, then aggregates the sentence scores into one document percentage (Turnitin FAQs).

The overlap matters more than it sounds. Because segments overlap, most sentences receive several scores that get pooled into one — which is why Turnitin warns that in a document of only a few hundred words the prediction is “mostly all or nothing”, with no opportunity to overlap. Short submissions get blunter verdicts, and a mixed human-and-AI passage in a short document can be flagged as entirely AI.

Turnitin’s own documentation makes a point that contradicts most explainers written about it, including older ones on vendor sites. Turnitin says its model is not programmed to evaluate perplexity or burstiness. Its current system is a deep-learning model based on the transformer architecture, trained on Turnitin’s in-house corpus of academic writing, and the company states its outputs come from “many learned patterns working together rather than by a small set of transparent, human-readable rules” — with the consequence that “individual predictions may not always be explainable in simple feature-by-feature terms”. If you want the interpretable version of the mechanism, our detector mechanism guide covers perplexity and burstiness properly; just do not assume Turnitin runs them.

In July 2026 Turnitin consolidated what had been a multi-model ensemble into a single model, which it says improves and simplifies the AI writing report while holding the sub-1% false-positive rate. That is the current architecture as of 30 August 2026.

One architectural point matters for anyone comparing Turnitin’s four language models: they are separate models with separate coverage. Turnitin’s Arabic model, released on 18 August 2026, is explicitly “a separate model from our English AI writing, Spanish AI writing, and Japanese AI writing models”, with its own list of detectable LLMs (Turnitin release notes). Every accuracy figure Turnitin publishes, and every study cited on this page, concerns the English model.

What Turnitin will and will not scan

The detector ignores far more of a typical document than most people assume, and the exclusions explain most confused reports.

Source: Turnitin file requirements and Turnitin FAQs. Note that Turnitin’s own pages disagree on languages: the FAQ page, updated 28 August 2026, lists Modern Standard Arabic alongside English, Spanish and Japanese, while the file requirements page still lists only the first three — Arabic support shipped on 18 August 2026, so the file requirements page is simply out of date. AI paraphrasing and bypasser detection remains English-only.

Only long-form prose counts towards the percentage. Turnitin does not analyse bullet points, lists, other short non-sentence structures, code, or bibliographies — references were excluded from processing in an August 2023 fix. This is why the percentage often does not match the amount of highlighted text: a document that is half bullet points is being scored on the other half. The denominator is “qualifying text”, not the submission.

Turnitin does not do code. The company states plainly that the model “does not reliably detect AI-generated text in the form of non-prose, or code”, and that it is “not pursuing ChatGPT code detection at this time”. AI detection is also not available in Gradescope.

Grammarly’s grammar checks are not flagged; Grammarly’s generative features are. Turnitin says its detector is not tuned to target spelling, grammar and punctuation corrections, and that in testing those changes were not flagged in most cases. It explicitly excludes Grammarly’s draft generation, paraphrasing and summarising features, which “will likely be flagged as AI-generated”.

How to read a Turnitin AI score

The indicator has four states, and only one of them is a number.

The asterisk is the most important thing on the indicator, and it is a self-admission. Since a release on 16 July 2024, Turnitin has attributed no score and no highlights in the 1–19% band, stating that “to avoid potential incidence of false positives, no score or highlights are attributed for AI detection scores in the 1% to 19% range” (Turnitin release notes). The company considers its own low-band output too noisy to print. Any institutional policy that treats an asterisk as evidence of anything is using a number Turnitin declined to publish.

The score deliberately undercounts. Turnitin states that in order to hold the false-positive rate at 1%, “there is a chance that we might miss some AI written text” — and gives the example that a document scored at 50% “could contain as much as 65% AI writing”. The detector is tuned to be safe rather than complete, which is the correct trade-off for a tool used in disciplinary settings and the direct cause of the false-negative results in the section below.

The report used to break the score into two categories. Since 4 August 2026 it shows one. From 16 July 2024, Turnitin separated text likely generated by an LLM from text likely AI-generated and then run through an AI paraphrasing tool or word spinner, with the second category highlighted in purple. A release on 27 August 2025 folded detection of AI bypasser tools — the products sold as humanisers — into the AI-generated category. Then on 4 August 2026 Turnitin merged the two categories entirely: “purple highlights previously used for AI-paraphrased text will no longer be shown”, while “the model will continue to detect likely AI generated content that may have been further modified by AI paraphrasers or bypassers” (Turnitin release notes). A single blue percentage now covers all three cases, and a report generated before 4 August 2026 will look different from one generated today on the same document.

Turnitin declines to name which bypassers it detects, on the stated grounds that publishing the list would help students evade it. Whether it works tool-by-tool is unverified: no independent test of Turnitin’s bypasser detection against named AI humanisers has been published.

Which AI models Turnitin says it can detect

As of its FAQ update on 28 August 2026, Turnitin’s English-language model lists detection coverage for output from more than 30 named large language models, including:

Source: Turnitin’s AI writing detection capabilities FAQs, updated 28 August 2026. Turnitin adds that it also detects “tools based on these LLMs”.

Two caveats belong next to that list. First, the same FAQ page prints conflicting release dates for the same models in two different sections — Claude Sonnet 4.6 appears as both February and May 2026, and o1-mini as both September and December 2024 — so treat the list as a coverage claim rather than a precise changelog. Second, a model appearing on the list is a claim of training coverage, not a measured detection rate for that model. Turnitin publishes no per-model accuracy figures.

What Turnitin claims about accuracy

Turnitin’s headline claim is a false-positive rate under 1% for documents where more than 20% of the text is flagged as AI. In its own wording: “we might flag a human-written document as AI-written for one out of every 100 fully-human written documents” (Turnitin FAQs).

The validation method is unusually specific for this market, and worth crediting: before every model update or release, Turnitin tests against more than 700,000 additional academic papers written before ChatGPT’s release — documents that cannot contain AI text — and adjusts the model to hold the rate under 1%.

The widely quoted “98% accuracy” figure is Turnitin’s, and it carries the same fine print: it applies to documents more than 20% AI-written. Below that threshold Turnitin does not publish a percentage at all.

Turnitin’s own caveats are more candid than most vendors’:

• “Turnitin does not make a determination of misconduct; rather, it provides data for the educators to make an informed decision.”
• The percentage “should not be used as the sole basis for action or a definitive grading measure by instructors.”
• False positives are more likely in “content without a lot of structural variation, text that literally repeats itself, or text that has been paraphrased without developing new ideas.”
• Documents under a few hundred words produce “all or nothing” predictions.

What independent testing actually found

This is where the picture in 2026 diverges from both the marketing and the folklore, and it diverges in a direction that surprises most people: the strongest recent evidence against Turnitin is about false negatives, not false positives.

The 2023 Walters study is the source of Turnitin’s best independent result, and it is three years old. William H. Walters compared 16 detectors and found that “two of the 16 detectors, Copyleaks and TurnItIn, correctly identified the AI- or human-generated status of all 126 documents, with no incorrect or uncertain responses”. Originality.ai came third, misclassifying two documents; overall accuracy across the remaining 13 detectors ranged from 63% to 88%. Walters himself flagged the limitation: “it is possible that TurnItIn performs especially well with the human-generated papers used in this particular analysis.” That study is still the one cited when Turnitin is described as the most accurate detector, including in Nature’s July 2026 feature on university detection (Nature 655, 535–537), which lists it as a reference.

The 2026 peer-reviewed study is the one that should change institutional thinking. Marijke Van Vlasselaer, Filip Van Droogenbroeck and Bram Spruyt ran a preregistered test of GPTZero, Copyleaks, Turnitin and Pangram against 160 English-language academic papers of at least 4,000 words each, in four categories. Their finding on Turnitin is stark: “Turnitin classified 100% of the Fully AI generated papers as False Negatives”, “consistently reported 0% for this category”, and — with GPTZero and Copyleaks — “completely failed to detect the AI-generated content”. On hybrid human-and-AI documents Turnitin reached 60.0% accuracy, and on humanised text 50.0%. Pangram, by contrast, reached 92.5% strict accuracy on both hybrid and humanised categories.

Two honest qualifications on that result. The AI text was generated in May 2025 — the fully AI papers with ChatGPT Deep Research, the inserted passages in the hybrid papers with GPT-4o — and Turnitin has shipped detection model updates on 14 October 2025, 12 February 2026 and 5 May 2026 since, plus the July 2026 architecture consolidation (Turnitin AI writing detection model release notes). The same documents resubmitted today may score differently, because Turnitin does not retroactively rescore. And the study’s design used long documents, which is the format Turnitin’s overlapping-segment method should handle best, so the failure cannot be explained away as a length artefact.

On false positives, the same study is good news for Turnitin. All four tools “classified 100% of human texts correctly”, and its authors concluded that “false positives were rare across all tools, suggesting improvement compared to earlier studies”. A second phase applied the best-performing tool to 1,163 real master’s theses from the 2024–25 academic year and flagged 45.5% of them, typically at low to moderate levels.

The reconciliation is straightforward and it is the same trade-off Turnitin describes in its own documentation. A detector tuned hard against false accusations will let AI text through. Turnitin has chosen that setting deliberately. The 2026 evidence is that against a current frontier model, on long-form academic writing, the cost of that setting is close to total.

The false-positive problem, and who it falls on

False positives are rarer than the internet believes and more consequential than the vendor’s 1% suggests, because they are not distributed evenly.

The most-quoted statistic on this subject does not apply to Turnitin. The 2023 Stanford study by Weixin Liang and colleagues, published in Patterns, recorded a mean false-positive rate of 61.3% on TOEFL essays by non-native English writers. It tested seven detectors: Originality.AI, Quill.org, Sapling, OpenAI’s classifier, Crossplag, GPTZero and ZeroGPT (arXiv preprint). Turnitin was not among them. Any page attributing that 61.3% figure to Turnitin — and there are many — is wrong. The underlying concern about second-language writers is real and is covered in how AI detectors work; the number does not belong to this tool.

Turnitin says it addressed that risk in training, stating it deliberately included “statistically under-represented groups like second-language learners, English users from non-English speaking countries, students at colleges and universities with diverse enrollments and less common subject areas such as anthropology, geology, sociology”.

The concrete evidence of harm is legal and regulatory rather than statistical, and one case is now on the record in detail. On 28 January 2026, in Matter of Newby v Adelphi Univ. (2026 NY Slip Op 26021), Justice Randy Sue Marber of the New York Supreme Court, Nassau County, granted an Article 78 petition brought by Adelphi University student Orion Newby, annulled the academic integrity finding against him, and ordered the university to expunge his record and rescind the sanction (full opinion). Newby, who was enrolled in Adelphi’s Bridges programme for students on the autism spectrum, submitted a World Civilization essay written with help from a programme tutor. His professor ran it through Turnitin, obtained what the court records as an “AI-generated score of 100%”, and filed a violation report alleging use of Grammarly — the one tool Turnitin’s own documentation says its detector is not tuned to flag for grammar and punctuation changes.

Three details from the opinion should be read by anyone writing institutional policy. The professor subsequently emailed the student to say he had assumed the integrity office would make its own determination, wording the court found “undermines both the strength and substance of the plagiarism claims”. The student submitted results from two other detectors indicating the essay was human-written, and the appeal officer “failed to even consider” them. And the same administrator who issued the determination also decided the appeal, a structure the court said “would deliberately thwart a student’s right to an avenue of meaningful ‘appeal’”. The court held the violation and the appeal denial “to be without valid basis and devoid of reason”.

Note that some coverage, including Inside Higher Ed on 11 February 2026, described the decision as a federal ruling. The opinion is a New York state court decision in an Article 78 proceeding.

In the UK, a July 2026 HEPI analysis of Office of the Independent Adjudicator case summaries published in July 2025 found that three of four AI-detection cases involved international or second-language students, and that in one case the OIA found the university had never considered whether AI detection performs less reliably for non-native English speakers.

The pattern in both jurisdictions is not that the detector is wrong often. It is that when it is wrong, the process around it fails the people least able to contest it.

Universities that have switched Turnitin’s AI detector off

The institutions closest to the tool have been walking away from it for three years, and the movement accelerated through 2025 and into 2026.

Two institution-level positions are worth separating from outright disabling. Yale University’s Poorvu Center does not endorse AI detection software or enable the features in Canvas. And on 13 August 2026, MIT’s Ad Hoc Committee on AI Use in Teaching, Learning, and Research Training recommended against relying on AI detectors, noting that MIT’s Committee on Discipline “does not consider AI detector output alone sufficient” — though that is a recommendation rather than Institute policy, and it does not name Turnitin.

Counts circulating online are unreliable. The most defensible published figure is Inside Higher Ed’s report of 5 August 2026 that “at least a dozen more” universities, beyond the named institutions that forbid treating detector output as proof, have disabled Turnitin’s AI detection software. Claims of “more than 50 universities” trace to SEO blogs with no stated methodology and should not be repeated.

Turnitin Clarity: the pivot from detection to process

Turnitin’s own answer to the accuracy problem is to stop relying on the score.

Turnitin Clarity is a composition workspace that records how a document was written rather than guessing whether a machine touched it. Announced at SXSW EDU on 4 March 2025 and available as a paid add-on to Turnitin Feedback Studio from Q3 2025, it gives educators visibility of “pasted text, typing patterns, construction time and draft history”, replayable as a video timeline, plus an optional AI assistant that an instructor can enable and that “refuses to write for” the student. Turnitin has not published list pricing; AI writing detection inside Clarity still requires the Originality add-on.

The evidentiary logic is better than a classifier’s. AI-pasted text arrives as a large block with little subsequent editing; genuine drafting shows incremental construction over sessions. That is observable behaviour rather than statistical inference, and it is the same shift towards process transparency visible across the category.

Turnitin’s first published Clarity data, released 24 February 2026, is also the source of the most-quoted number about student AI use: 14.8% of English-language submissions between October 2025 and February 2026 contained 80% or more AI-generated writing, up from 3.3% between April and August 2023 (Turnitin press release). Read that figure with the false-negative evidence above in mind: it is a floor, measured by a detector tuned to undercount.

The same release reported that 29% of student prompts in Clarity asked for review, judgement or feedback; 94% of students wrote their own prompts rather than using suggestions; and only 36% of feedback-category prompts were judged effective. The behavioural picture is students seeking help with their own work far more often than commissioning replacements for it.

What to do if Turnitin flags your work

If you are a student who has been flagged, your strongest evidence is your writing process, not a counter-test. Running the essay through GPTZero or another detector produces a second opinion from a different black box, and the Adelphi case shows courts and appeals panels weigh process evidence more heavily. Preserve version history in Google Docs or Word, drafts, notes, browser history, outlines and the timestamps on all of them. Ask which tool produced the score, what the exact percentage was, and whether it was above or below 20% — because below 20% Turnitin itself does not publish a number. Ask whether the institution’s policy permits detector output as sole evidence; most now say it does not.

If you use AI legitimately, the defence is documentation, not avoidance. Turnitin’s mark records that a model touched the prose, not what it contributed — so disclose the tool, keep the prompts, and keep the drafts that show your own reasoning between them. The same applies to the assistants covered in best AI for students: permitted use that is documented survives a flag, undocumented use of any kind does not.

If you write in English as a second language, or you have a diagnosed learning difference, say so early and in writing. Both the OIA cases and the Adelphi lawsuit turned on institutions failing to consider circumstances that plausibly explain a polished or unusual style. That is a procedural failure you can name in an appeal.

If you are an instructor, Turnitin’s own guidance is the defensible starting point. Its help documentation tells instructors to refer to institutional policy first, treat the result as formative rather than punitive, and use the percentage as “a good starting point for discussing” the work — with the final determination resting with the instructor, not the tool (Turnitin guidance). Its scripted conversation openers ask the student to explain their process and review the report together, rather than opening with an accusation.

If you are setting institutional policy, the 2026 evidence supports three rules. Never allow a detector score as sole evidence. Never treat an asterisk as a finding. And build an appeal that actually considers the student’s evidence, because the one court to examine the question found that failing to do so voided the finding.

Where this leaves Turnitin

Turnitin’s position in 2026 is genuinely awkward, and it is not the position its critics usually describe. Its false-positive discipline is real and unusually well documented — the 700,000-paper validation set, the withheld 1–19% band, the deliberate undercounting, the repeated statement that its own output is not proof. On the measure that generates the headlines, it is more careful than most of the market.

The cost of that discipline is now measurable. Against a 2025-generation frontier model on long academic documents, peer-reviewed testing found the detector returning zero. A tool that will not accuse the innocent and cannot catch the guilty is not a detector in any useful sense; it is a conversation starter with a percentage attached. Turnitin appears to know this, which is why its product roadmap in 2025 and 2026 runs through Clarity and writing-process visibility rather than better classification.

The broader direction of travel points the same way. Provider-side AI watermarking went mainstream in August 2026 with Anthropic’s global text marking and Google’s SynthID, offering real evidence rather than inference — for the fraction of AI text that carries a mark. Between watermarks at the source and process telemetry at the desk, style-guessing classifiers are being squeezed from both ends. Turnitin’s advantage was never the model anyway; it was being installed at roughly 16,000 institutions in 185 countries. Whether that distribution survives the accuracy question is the open item for the 2026–27 academic year.

Frequently asked questions

Does Turnitin detect AI writing?

Yes, for institutions that license Turnitin Originality or the iThenticate 2.0 AI writing add-on. Turnitin adds an AI writing indicator to the Similarity Report showing what percentage of a submission’s long-form prose it predicts was generated by an AI writing tool, or generated and then modified by an AI paraphraser or bypasser. The feature is off for institutions that have not licensed or have disabled it, and Turnitin does not process submissions for AI writing in that case.

How accurate is Turnitin’s AI detection in 2026?

Turnitin claims 98% accuracy with a false-positive rate under 1%, but both figures apply only to documents where more than 20% of the text is flagged, and Turnitin validates the false-positive rate against 700,000 academic papers written before ChatGPT’s release. Independent findings diverge sharply by date: a 2023 study by William H. Walters found Turnitin correctly classified all 126 test documents, while a peer-reviewed study published in the International Journal for Educational Integrity in June 2026 found Turnitin scored 0% on 100% of fully AI-generated papers while producing no false positives on human papers. The honest summary is that Turnitin is tuned heavily against false accusations, and pays for it in missed AI text.

Can Turnitin detect ChatGPT, Claude and Gemini?

Turnitin’s published model list includes output from GPT-4o through GPT-5.4, Claude 3 Haiku through Claude Sonnet 4.6 and Claude Opus 4.5, and Gemini 1.0 Pro through Gemini 3.1 Pro preview, along with Llama, Mistral, DeepSeek, Grok and Amazon Nova models. That list is a claim of training coverage, not a measured per-model detection rate — Turnitin publishes no accuracy figures broken down by model, and 2026 peer-reviewed testing found it failed to flag long documents generated with a current OpenAI research model.

What does the asterisk mean on a Turnitin AI score?

An asterisk means Turnitin detected AI writing somewhere between 1% and 19% of the qualifying text and has deliberately withheld the exact number and the highlights. The policy took effect on 16 July 2024, and Turnitin’s stated reason is to avoid false positives, since it considers scores in that band too unreliable to attribute. An asterisk is therefore not a finding, and no institution should treat it as one.

Can students check their work in Turnitin’s AI detector before submitting?

No. Turnitin states that the AI writing indicator and report “are not visible to students”, and there is no consumer version of the tool — it is sold to institutions only. An instructor can download the AI report as a PDF and share it with a student, but students cannot run a pre-check. Services advertising themselves online as Turnitin AI checkers for students are not operated by Turnitin.

Does Turnitin detect AI humanisers and paraphrasing tools?

Yes, in principle. Turnitin added AI paraphrasing detection on 6 December 2023 and extended it to AI bypasser tools on 27 August 2025, so the AI writing report now estimates what share of the flagged text was run through a humaniser. Turnitin declines to name which tools it detects, on the grounds that publishing the list would help students evade it, and no independent tool-by-tool test of that capability has been published. The 2026 International Journal for Educational Integrity study measured Turnitin at 50.0% accuracy on humanised text.

How many words does Turnitin need to generate an AI score?

At least 300 words of prose text in a long-form writing format, and no more than 30,000 words, in a file under 100 MB and in .docx, .pdf, .txt or .rtf format. Only prose sentences inside paragraphs count — bullet points, lists, code and bibliographies are excluded from the score entirely, which is why the percentage often does not match the amount of highlighted text. Turnitin also warns that documents of only a few hundred words produce near “all or nothing” predictions.

Which universities have turned off Turnitin’s AI detector?

Vanderbilt disabled it in August 2023, Georgetown in October 2023 and Johns Hopkins in September 2023; UCLA opted out of the preview feature in April 2023. More recently, the University of Queensland removed the AI Writing Indicator for Semester 2 2025, the University of Waterloo discontinued it from September 2025, the University of Cape Town from 1 October 2025, and Curtin University from 1 January 2026. Inside Higher Ed reported on 5 August 2026 that at least a dozen universities have disabled the software; higher counts circulating online do not trace to a stated methodology.

What should I do if Turnitin falsely flags my essay as AI?

Preserve your writing process evidence first: document version history, drafts, notes, outlines and timestamps carry more weight than a second detector’s opinion. Ask for the exact score and whether it was above or below the 20% threshold, since Turnitin withholds numbers below it. Check whether your institution’s policy allows detector output as sole evidence — most now say it does not, and Turnitin itself states its results should not be the sole basis for action. If English is not your first language, or you have a diagnosed learning difference, put that in your appeal in writing: on 28 January 2026 a New York court annulled an Adelphi University academic integrity finding as “without valid basis and devoid of reason” after the university failed to consider a student’s evidence and let the same administrator decide both the case and the appeal.

Is Turnitin’s AI score the same as the similarity score?

No. Turnitin states the two are completely independent and do not influence each other. The similarity score measures matching text against Turnitin’s content database and is a plagiarism signal. The AI writing percentage is a prediction about whether a language model generated the prose, with no database involved. A document can score high on one and zero on the other, and conflating them is one of the most common errors in student-facing advice about Turnitin.

Written 30 August 2026 from Turnitin’s own documentation as updated 28 August 2026, plus peer-reviewed and primary institutional sources. Turnitin’s detection model has been updated at least four times since October 2025 and does not rescore past submissions, so independent test results are dated to the model generation they measured. Vendor accuracy figures are reported as claims and kept separate from independent findings throughout.

← All guides