Turnitin checker, scope and limits

Authorship is not a property of a text. Nothing in a document records who or what produced it, so no analysis of the words alone can recover that fact. What an analysis can recover is distributional: how much sentence length varies, how often constructions repeat, how evenly the argument is scaffolded, how predictable the next clause is given the last one.

Every AI detector, including the one institutions buy, is inferring from those distributional features to a conclusion the features do not contain. That inference can be well calibrated or badly calibrated, but it remains an inference. This checker performs the same kind of reading and reports the features rather than pretending the conclusion is available.

Stated first because the category makes it necessary. We hold no affiliation with Turnitin, no licence, no access to their model, and no visibility into any report they generate. Their detector is licensed to institutions and its internals are unpublished.

Two things follow. Any site claiming to replicate that output is claiming access to a proprietary system it does not have, and any correspondence between an independent tool's reading and an institutional report is unmeasured rather than demonstrated. A search for this term returns a row of domains assembled from the trademark itself, most of which never state their relationship to it.

The constraints are worth internalising before the first run, since the allowance is small enough that method matters.

Step 1

Select a coherent unit of text. The bound is 1 to 12,000 characters and it is enforced before the model call, so an oversized submission fails at the boundary rather than returning a truncated reading.

Step 2

Submit. The text is transmitted to our server because analysis requires a model to read it. This route offers no local mode, and describing it as private would be false.

Step 3

Treat the findings as hypotheses about your own prose and test each against what you remember writing. The ones you can account for are noise. The ones you cannot are the finding.

The endpoint returns a single string of analysis. There is no score field, no band classification and no per-sentence structure anywhere in the implementation, so none is offered. Competitors in this result set publish all three, along with accuracy figures and false-positive rates presented to one decimal place.

A percentage implies a calibrated model with a measured error rate against a known ground truth. Absent that measurement, the number is a rendering choice. We would rather publish the reasoning than a figure whose only function is to look quantitative, and the analysis is additionally constrained: it is instructed never to claim certainty about authorship, plagiarism or which model wrote a text, and to describe observable signals and limitations instead.

In this setting the two error types are not symmetrical. A missed detection costs an institution a marginal enforcement failure. A false positive costs a student an accusation they cannot disprove, since the absence of evidence of machine authorship is exactly what an honest draft looks like.

The writing that draws false positives is identifiable and consistent across the sources examined: formal academic register, prose written tightly to a rubric, heavily templated structure, and the phrasing patterns of writers working in a second language. These are all features of disciplined student work, which means the error concentrates on the population least able to absorb it.

Long writing contains flat passages by necessity. Method sections, definitions and transitions are mechanical because their function is mechanical, so a count of mechanical sentences across a long document is close to uninformative.

Concentration is the informative variable. A single paragraph in which every sentence is uniform in length, predictable in structure and generic in phrasing reads differently from the same number of such sentences distributed over twenty pages, and it is what a human reader notices once a report has been opened. Read the analysis for clusters and treat the scattered observations as the background rate of ordinary prose.

Turnitin's older and better-established function is similarity detection against published sources and an archive of previously submitted student work. That capability requires a corpus. We operate none, perform no matching, and produce no similarity figure.

This is a real limitation rather than a positioning statement. An uncited paraphrase, a quotation whose attribution was lost in editing, or reused text from a previous submission will not be surfaced here by any means. Those require a similarity service, and this route is not one.

Unknowable, and therefore not claimed. Their output is not observable to us, so any statement about the effect of this tool on it would be fabricated rather than measured.

The defensible claim is narrower and still worth something. Prose that is uniform, repetitive and generic is weak on its own terms, independent of any detector, and an analysis that identifies where yours is weak lets you fix it and explain it. A student who understands their own draft is in a stronger position than one holding a rewritten file and a number from an anonymous domain.