CiteOwl

Blog/Using AI honestly

Can Turnitin detect Gemini? Seven versions and one gap

The CiteOwl Team· Updated 20 September 2026· 12 min read

Yes, for the seven Gemini builds Turnitin names on its published list, the newest of them a preview from February 2026. Four Gemini releases since then are missing. The app hands you a model called Fast rather than a version number, and Google's own mark on Gemini text is not one your department can read. Write from sources you can open, and put a quote under every claim.

Turnitin names seven Gemini versions

Turnitin's AI writing detection is a separate feature from the similarity score, and it does not compare your essay against a library of chatbot output. Sentences are pulled from your submission, segmented into overlapping sections, and each segment is scored between 0 and 1 for how likely it is to be human or machine written, with the pooled scores producing the percentage an instructor sees. It is the same classifier behind a ChatGPT score, and the report never names a model.

What Turnitin does publish, unusually for a detection vendor, is a list of the specific models its English detector picks up, each with a release date. Reading it on 20 September 2026, the list runs to 32 entries and seven of them are Gemini: Gemini 1.0 Pro, Gemini 1.5 Pro, Gemini 2.5 Pro, Gemini 2.5 Flash, Gemini 3 Flash Preview, Gemini 3 Pro Preview and Gemini 3.1 Pro Preview. The sentence closes with a clause worth reading twice. Coverage extends to "tools based on these LLMs as well", which reaches the study apps and writing assistants running on Gemini underneath without saying so.

Turnitin is also blunt about how the score is produced. Its model, it says, "is not explicitly programmed to evaluate specific signals such as 'burstiness,' 'perplexity,' or other individual metrics sometimes referenced in public discussions". The advice circulating about Gemini's sentence rhythm is arguing about a dial Turnitin says it does not turn.

Four releases since, and no Gemini 2.0 at all

The newest Gemini on that list is Gemini 3.1 Pro Preview, which Google's own Gemini API release notes date to 19 February 2026. Five generally available Gemini models have shipped since, the four Flash releases in the table and a 3.5 Flash-Lite in July, and none of them appears.

The list is patchy behind its own edge too. Gemini 2.0 Flash is on none of it, although Google's deprecation table dates its release to 5 February 2025 and its shutdown to 1 June 2026, sixteen months in which it was the Gemini most students met. And one row contradicts Google outright: Turnitin dates Gemini 2.5 Flash to March 2024, while Google's own table gives 17 June 2025. Turnitin's date puts 2.5 Flash a year before the 2.5 Pro sitting beside it in the same sentence. That is a hand-maintained document, not a live registry, and we found the same thing on the Claude rows when we went through what the list says about Claude.

Here is where the humanizer sites want you to draw a conclusion, so be careful. An absence from that list is not a statement that a model gets through. Turnitin does not claim its detector generalises to models it has not named, and it does not claim the opposite either. It does say the classifier learns statistical patterns rather than model identity, and that in July 2026 it consolidated a multi-model ensemble into a single model while holding its false positive rate under 1%. A detector built that way is not promised to fail on an unlisted model. Nobody, Turnitin included, publishes a number for the version you used.

One practical note follows from the same page. Model updates never rescore old work. A submission is only re-examined if somebody submits it again.

The app never told you which Gemini you used

Suppose the list were complete and current. You still could not use it, because the thing you typed into does not print a version number on its answer.

Google's own announcement of Gemini 3 Flash tells you how the app presents its models: Gemini 3 Flash "can be accessed in the Gemini app by selecting 'Fast'", with Gemini 3 Pro available as "Pro" in the model picker. Fast and Pro are what you choose. Which build answers behind those words moves whenever Google ships, and Google has shipped four times since February.

The plan matters as well. Google's subscription page lists the free plan as reaching Gemini 3.6 Flash, the model its release notes date to 21 July 2026, with varying access to 3.1 Pro. So the everyday Gemini a student on the free plan writes with is a model Turnitin has not named, while the one they occasionally get bumped up to is on the list as a February preview. Plenty of students have both: in a post dated 19 August 2026 Google offered eligible college students a free year of a paid AI plan, redeemable to 31 December 2026, which is the most effective distribution any of these assistants has had on a campus.

Put those together and the question loses its shape. You cannot look up your own essay on that list, because you do not know which row it belongs to, and the row it belonged to in July is not the row it would belong to now.

What the published tests actually measured

Two peer-reviewed studies have put Google's models through Turnitin. Neither is current, and both are worth more than anything else on this query.

The first is directly about Gemini. In the Journal of Applied Learning and Teaching in early 2025, Muhammad Abid Malik and Amjad Islam Amjad generated fifteen essays of up to 500 words, five each from the free versions of ChatGPT 3.5, Perplexity and Gemini, then made three further copies of each set: edited through Grammarly, paraphrased through QuillBot, and edited 10% to 20% by a human language expert. Sixty documents went through Turnitin, ZeroGPT, GPTZero and Writer AI. Turnitin returned a 100% AI score on every file, including every Gemini file, and held it through all three techniques. The other three detectors did not: QuillBot cut ZeroGPT's average on the Gemini essays by two thirds, from 95.4% to 31.8%.

Read the limits with the result. Five Gemini essays, 500 words each, one pass through each detector, and the free Gemini of 2024, generations before anything in the app today. Turnitin sat at 100% on the raw text, so nothing could show a difference in either direction. The authors' own conclusion is that a score "should not be taken as a final verdict, and further checks and investigations should be carried out before labelling a text as AI-generated or otherwise", and their paper notes an earlier test that put Turnitin at 0% on GPT-4 text.

The second study tested Bard, which is the same model line under the name Google used before the rename. A 2024 paper in the International Journal of Educational Technology in Higher Education generated text with GPT-4, Bard and Claude 2 and ran it past seven detectors. Bard was the most detectable generator in the set, with 76.9% of its outputs identified, and Turnitin caught 61% of AI text overall, second of the seven.

What happened next is the part people skip. Under adversarial editing, accuracy across detectors fell 17.4% on average, and the drop was wildly uneven by generator: Bard lost 38.8%, against 8% for Claude 2 and 7.6% for GPT-4. Turnitin itself took the largest hit of any tool tested, 42.1%, falling from second place to fifth. The most detectable model was also the most fragile once anyone touched its output, which is the opposite of the tidy story on both sides of this argument.

Nobody has published a test of Turnitin against any Gemini 3 model. That is the honest state of the evidence in September 2026, and any page telling you otherwise is about to show you a number it made up.

Google marks Gemini's text, and cannot check it for you

There is a second detection system in this story, and it belongs to Google rather than to your university. Google DeepMind's SynthID watermark, which most people know from images, now covers text: "We've expanded SynthID to watermarking and identifying text generated by the Gemini app and web experience." Nothing is added to the words. Large language models pick each next word from a set of candidates with probability scores, and SynthID adjusts those scores so the choices themselves carry a pattern.

Two things about it matter more than the mechanism. The first is that Google makes no durability claim for text. For images it says the mark is designed to survive cropping, filters, frame rate changes and lossy compression, and for audio that it survives noise, compression and speed changes. For text, that sentence simply is not there.

We went through the whole marking landscape, Google and OpenAI and Anthropic together, in does AI text have a hidden watermark, and the things you actually can inspect in a file are in how to check a document for hidden marks.

The percentages on this search have no method

Search this question and you are handed numbers to the decimal place. We opened the pages handing them out.

One states that Turnitin's accuracy "has reached up to 98-100% for standard AI text", with no source, no study and no method, on a site whose call to action is to hire its assignment writing service. Another reports that "Turnitin claims 91% accuracy on raw Gemini output" and that independent testing found Turnitin flags "about 61% of Gemini-generated passages" at a 20% threshold, naming no institution, no date and no paper, on a page selling a tool that promises to make Gemini text undetectable. It then prints per-variant rates, roughly 70% for 2.5 Flash and roughly 53% for 2.5 Pro, as though somebody had measured them.

The claim attributed to Turnitin is checkable, and it is wrong at the root. Turnitin's FAQ publishes no per-model accuracy number for any model on its list, Gemini included. What it publishes is a false positive rate, under 1%, and only for documents scored above 20% AI, a condition dropped every time the number is quoted. It also tells you it undercounts deliberately to keep that rate down, giving the example that a document read as 50% AI "could contain as much as 65% AI writing", and that before each model release it validates the rate against over 700,000 academic papers written before ChatGPT existed. A company telling you it misses AI writing on purpose is not one claiming 91% on anything.

The pattern is the same one we found on the QuillBot query, where a page reporting a controlled test failed its own arithmetic in four subgroups out of four. Precision and evidence run in opposite directions here. The page with a real method, printed in a journal, reports a single crude figure and spends a paragraph on what it cannot support.

The answer that survives the next release

Both systems in this article are weakest at the case you are probably in. Most people asking whether Turnitin can detect Gemini did not paste an essay out of the app. They drafted two paragraphs with it, or asked it to tidy a section, or had it summarise three papers and wrote from the summary.

Turnitin suppresses scores from 1% to 19% entirely, showing an asterisk instead, because false positives are likeliest there. On a short piece it warns that the prediction "will be mostly 'all or nothing'", so a genuine mix can come back as entirely machine written. Google's mark is written into the choices Gemini makes between candidate words, so by construction there is nothing for it to sit in where the choosing was yours. Neither system separates the essay you wrote from the essay you were helped with, and that is the only distinction your department is asking about. Our guide to what a Turnitin AI score actually means goes further into reading one.

What does not move is the work underneath. Gemini invents references like every model does, and a fabricated citation stays wrong through any number of releases and any amount of rewriting. A marker who clicks one dead DOI has something far more concrete than a score, and anyone can run that check in a minute with our free citation checker. We cover the failure itself in does Gemini make up citations.

So ask the question that has an answer in a year's time. Not whether your model is on a list, but whether you can stand behind every sentence and open every source. CiteOwl is built for that and nothing else: its agent researches real papers, attaches each factual claim to a source with the quote that supports it, and proposes every change as a diff you accept, reject or edit. Get that right and the list becomes trivia. If you are choosing between assistants on research quality rather than detectability, our comparison of Gemini and ChatGPT for research papers is the more useful question.