Gemini Originality checker signals

The input is a string of 1 to 12,000 characters. The analysis is performed by a language model reading that string and nothing else, and the output is its written account of what the prose is doing: sentence length distribution, the recurrence of connectives, hedging density, where the register shifts and where it refuses to.

That is the entire scope, and the category around it is vague on purpose. This route holds no index, queries no search engine and consults no archive of submitted work. A finding here is a description of your text, derived from your text.

The search term collapses two operations that share no mechanism. Corpus matching takes your text, compares it against a body of documents, and returns overlap with locations. Pattern analysis takes your text, compares it against nothing, and returns a characterisation of its style. The first is retrieval. The second is reading.

The market reflects the split without naming it. Originality.ai runs retrieval, with a hybrid index: public web content through Google search for web-app users, extended to academic journal databases and private institutional repositories for LMS customers. DetectGemini runs reading, and is candid about it, describing its method as patterns, perplexity and burstiness with no comparison corpus involved.

This route is the second kind. The consequence follows directly: a result from here can never be evidence that a passage is or is not copied, and no wording on this page will imply otherwise.

The output is prose, not a dashboard, which means reading it takes a little more care than glancing at a number.

Step 1. Submit a coherent unit of text

One section, one argument, within the 12,000-character ceiling. Fragmented input produces fragmented findings, and a passage under roughly 50 words carries too little variation for any conclusion, a threshold DetectGemini publishes for its own tool.

Step 2. Constrain the question

The instruction field accepts 300 characters. An unconstrained read returns a general impression; a constrained one returns something testable. Ask about paragraph openings, or clause length, or whether the vocabulary holds a consistent register.

Step 3. Verify each finding against the text

Every claim in the output points at something you can go and look at. Check it. The findings that survive inspection are the ones worth acting on, and the ones that do not are why this page publishes no score.

Google has built provenance infrastructure and documents it publicly. SynthID embeds invisible watermarks identifying images, video and audio generated or edited by Google's models, robust through rescaling, recolouring and compression. Content Credentials, the Coalition for Content Provenance and Authenticity standard, attaches origin and history to a file across devices and applications.

Neither covers text. Google's own help page limits SynthID to images, video and audio, and notes that SynthID recognises Google's tools rather than every model. It also states that Content Credentials do not tell you whether a file is definitely AI-generated, only where it came from.

That is the structural fact underneath this entire search term. There is no signature in a paragraph of Gemini output to find. Every product ranking here, at every claimed accuracy, is inferring from style. The difference between them is not method, it is how openly they say so.

Four figures appear across the pages ranking for this term, and almost none arrives with a test set, a base rate or a definition of a positive. One vendor does disaggregate, and its own numbers argue against quoting any single figure.

Up to 99.5 percent

Originality.ai on global plagiarism, meaning wholesale reproduction of an existing document. The easiest case retrieval can be given.

67.5 percent

The same tool on patchwork plagiarism, with named competitor figures alongside it: Copyscape 30, Grammarly 51, Turnitin 52.

41 percent

The same tool again, on paraphrased plagiarism. The case that matters most in practice, and the one where retrieval fails more often than it succeeds.

99 percent

DetectGemini's figure for pattern detection, stated repeatedly and without a corpus, a methodology or a false positive rate anywhere on the page.

DetectGemini lists among its own limitations that heavily edited or paraphrased AI content may evade detection. Originality.ai measures its paraphrase accuracy at 41 percent. TDMP, testing ChatGPT, Perplexity and Gemini against three plagiarism detectors, found that when the models were asked directly to rephrase a passage, all three produced text the detectors judged original.

Retrieval and reading fail at the same point for different reasons. Retrieval fails because paraphrase destroys the string overlap it depends on. Reading fails because an edit pass changes exactly the surface statistics it measures. A single intervention defeats both, which bounds what a negative result from either can mean.

The correct reading of a clean result is therefore narrow: this pass raised nothing. It is not a finding of originality, and nobody on this search can produce one.

Everything below is a property of how this route is built, taken from the code rather than from positioning.

No corpus, no source matching

No index, crawler or search step exists on this route. It cannot report whether a sentence appears elsewhere, and that is the largest capability gap against the retrieval tools in this market.

No score, no verdict

The engine is instructed never to claim certainty about authorship, plagiarism or which model produced a text, and to describe observable signals and limitations instead. The output reflects that instruction.

Text is sent to our server

This route is remote. Your input is transmitted to our AI server for the analysis, which is stated in the widget above the run button. There is no local-only mode.

5 runs per rolling 24 hours

A hard allowance enforced per requester. The window rolls, so the oldest run expires rather than the count resetting at a fixed time.

Non-deterministic output

A language model performs the read, so wording varies between runs on identical input. The substance of the findings should be stable; the phrasing will not be.

No, and the honest vendors concede it in their own limitations sections. DetectGemini, under a headline claim of 99 percent, acknowledges that some human writing styles naturally resemble AI patterns and that no detector is 100 percent accurate.

The asymmetry that matters is in the consequences. A false negative costs a reviewer nothing. A false positive costs a specific person an accusation they cannot disprove with another test, because a second stylometric read is not independent evidence, it is the same inference run again.

So this page publishes no percentage and no verdict, and the tool names no model as the author of anything. If you are on the receiving end of an accusation, the material that carries weight is process: drafts, edit timestamps, research notes, the version that existed before the one being questioned.