Claude Turnitin checker without a score

A written account of the properties in a submitted passage that pattern-based detection is built to measure: the distribution of sentence lengths, the predictability of the phrasing, the degree to which claims are qualified rather than asserted, and whether successive paragraphs are built to one template. Findings are quoted back as passages a person can revise.

The route is server-side. Input is transmitted, bounded at 1 to 12,000 characters, with an optional 300 character instruction. Five runs per rolling 24-hour window, no account. Analysis runs at temperature 0.2 rather than the 0.55 used on rewriting routes, and completion is capped at 1,400 tokens.

First, no score. Nothing in this path computes a number, so a percentage would be manufactured rather than measured. The system instruction governing the route forbids claiming certainty about authorship or plagiarism and requires observable signals and limitations to be described instead, which rules out the estimate as well as the assertion.

Second, no similarity. A plagiarism report is a retrieval operation: your text against an indexed repository of publications, web pages and prior submissions. There is no repository in this codebase. Without a corpus the operation is not merely unimplemented, it is unavailable, and a page that keeps the word plagiarism in its promise while disclaiming it in an FAQ is trading on the ambiguity.

We have no affiliation with Turnitin, submit nothing to them and retrieve nothing from them. That is a statement about architecture, not a policy we could relax.

Submit a unit long enough to have a distribution. Pattern measures are statistical over a stretch of text, so a single paragraph yields nothing actionable from any reader, this one included.

Step 1

Paste one section, 1 to 12,000 characters. Characters and words are metered live above the field.

Step 2

Declare the genre and level in the 300 character instruction field. A lab report and a reflective essay have different expected registers, and salience is relative to the expectation.

Step 3

Execute, then revise the named passages by adding material only you hold: your sources, your page numbers, your counterexample. Remaining runs are printed on the widget.

The pages competing on this term divide cleanly. One publishes a table of per-model detection rates derived from its own sample set and concludes, in the same article, that you should buy its humanizer. Another publishes no figure at all and writes instead that nobody can promise an outcome on your exact draft. The second is the more useful document.

A vendor-run rate is a measurement of that vendor's samples against a third party's product on the day they tested it, reported by a party with an interest in the result. It is not a property of your text. Treating it as a probability for your submission is a category error, and it is the error the entire humanizer category is built to encourage.

There is a further problem with any number here. Turnitin's own documentation on how the AI writing report behaves returned HTTP 403 when we tried to read it, so the widely repeated specifics about confidence bands and highlight colours are, from where we stand, third-hand. We do not restate them as fact.

One of the better third-party accounts describes the measurement as two quantities: how predictable each word choice is, and how much sentence length and complexity swing within a paragraph. When both stay low across enough text, the indicator leans toward likely AI.

Compressed length variance

Sentences clustering within a narrow band. The absence of swing is the signal, not the average length, so trimming words does not address it.

High-probability phrasing

Each next word being the expected one. Synonym substitution moves the wording without moving the predictability, which is why paraphrase tools underperform on this measure.

Qualified assertion

A claim followed by a softener. It flattens the argumentative contour, since nothing is ever risked and nothing is ever conceded.

Template paragraph structure

Identical internal architecture repeated across sections. Formally correct and statistically conspicuous at document scale.

Non-specific evidence

Support that would fit any essay on the topic. The fix is the only one that reliably works: material that could not have come from anywhere but your reading.

Detection is probabilistic, and the errors are not distributed evenly. The most careful page in this result set concedes real false positives and names the two populations they concentrate in: non-native English speakers, and writers of formal, formulaic prose. Both produce text that is low on variance for reasons unrelated to any model.

This matters for how our own findings should be read. This route measures the same properties, so it will describe that writing as flat, correctly, while saying nothing whatsoever about how it was produced. A report that a passage is low on variation is a fact about the passage. Any inference from that to authorship is the reader's, not one the tool supports.

No. The route is instructed never to claim certainty about authorship, plagiarism or which model produced a text, and the same limitation is recorded against this engine family in the capability contract in this repository. It is a constraint on the runtime, not a caption added to the page.

Nor do we predict what an institutional detector will return, tune anything against detector scoring, or describe any outcome as passing. What your institution permits is set by its policy, and no analysis from us alters it.

The result set for this term is composed almost entirely of two things: humanizer products whose articles conclude in their own checkout, and assignment-writing services publishing explainers as acquisition. Two near-name domains occupy the same slug and advertise unlimited free local processing.

Turnitin does not compete for this phrase at all. It owns the query about how its report works, and that documentation was unreachable to us. The practical consequence is that a reader researching this subject is mostly reading interested parties, which is a reason to prefer any page that tells you what it cannot do before it tells you what it can.