AI assignment checker

Two different questions get collapsed into one on almost every page in this category. The first is whether a passage exhibits features associated with machine-generated writing. That is answerable, approximately, from the text alone. The second is whether a particular person used a particular tool to produce it. That is not answerable from the text at all, at any confidence, by any method available to a reader of the finished string.

This tool answers the first question and refuses the second, and the refusal is in the code rather than in a disclaimer. The classifier that runs this route is instructed never to claim certainty about authorship, plagiarism, or which model wrote a text, and to describe observable signals and limitations instead. What returns is prose about the writing. There is no score.

They contradict each other, and the contradiction is the finding. Turnitin has claimed a false positive rate below 1%. A Washington Post test, on a smaller sample, reported figures as high as 50%. In June 2023 Turnitin acknowledged that its rate was higher than it had originally asserted. In the other direction, a University of San Diego law library guide reports that the same checker misses roughly 15% of AI-generated text, a tolerance accepted deliberately to hold the false positive rate down.

The vendor pages sit in the same result set without engaging any of it. GPTZero displays 99% and 99.9%, alongside 95.7% detection with 1% human misclassification on the third-party RAID benchmark. The same page states that results should not be used to punish or as the final verdict, and that no AI detector is 100% accurate. Nothing there is dishonest in isolation. Read together, the headline number and the caveat describe different products.

Ordered so the cheap deterministic work happens before any of the five runs is spent.

Step 1

Clear the deterministic layer first: run the hidden character scanner, then verify each citation exists and supports the claim attached to it.

Step 2

Submit the section you have least confidence in, up to 12,000 characters, with a narrowing instruction if quotations or data should be ignored.

Step 3

Treat the findings as a list of sentences to reconsider, and edit individually rather than re-running the same passage for a better result.

A percentage carries an implicit claim about calibration: that a document scoring 80 is meaningfully different from one scoring 60, and that both figures mean the same thing across genres, lengths and first languages. None of the published evidence supports that. A number also compresses an argument into a symbol, which is precisely what makes it circulate, get screenshotted and get used as proof.

Findings resist all of that. A sentence named as a formal connector stack is either a connector stack or it is not, and you can look. The output is checkable against the text in front of you rather than against a claim about a model you cannot inspect. That is a smaller promise and a more defensible one.

This is the part of the evidence that matters most to an individual reader, because an aggregate false positive rate says nothing about your personal risk if you are in an over-flagged group. Research cited by the University of San Diego guide, drawing on Stanford's Human-Centered AI institute and on Liang and colleagues in Patterns, reports that students writing English as a second language and neurodivergent students are flagged at higher rates than others.

Institutions have begun to act on this rather than only cite it. The University of Pittsburgh's teaching centre states that most such tools have a high false positive rate and are more likely to misfire on non-native English speakers, and it has disabled the AI detection tool in Turnitin while endorsing no detector at all. That is a published institutional position, and it is usable in a conversation about a score.

Under the probabilistic argument there is a small deterministic layer, and it is the only part of this subject with a definite answer. Text pasted out of a chat interface can carry invisible characters through the paste: zero-width spaces, a zero-width non-joiner, a word joiner, a stray byte-order mark, no-break and narrow no-break spaces that render identically to ordinary ones, directional controls.

The scanner on this site handles exactly 14, by codepoint, in the browser, with no model and no probability. Either a zero-width space is in your file or it is not. Clear that layer first. It costs seconds, it has no false positive rate, and it removes an artefact that is difficult to explain after the fact.

Stated as a list because the omissions define the tool. No percentage or confidence score. No identification of which model produced a text, which the system prompt forbids explicitly. No per-sentence highlighting, since the surface is two text panels and prose comes back. No plagiarism or source matching, because there is no corpus in this codebase. No prediction of what Turnitin, GPTZero or any institutional checker will report, and no affiliation with any of them. No accuracy figure of our own, because nothing here measures one.

The numbers that do apply: 1 to 12,000 characters per submission, output capped at 1,400 completion tokens, 5 runs per rolling 24 hours, a 300 character instruction field, and a sampling temperature of 0.2 on an analysis run, which is lower than the 0.55 used for rewriting and is why repeated runs vary less than they would elsewhere on this site. Paste only, with no file upload and no integration of any kind.

This page addresses one case: a person reading their own draft before submitting it. In that case the useful work is concrete. Clear the invisible characters. Confirm every source exists and supports what it is cited for, since a fabricated reference is a violation whether or not any detector notices it. Fix the specific sentences that read as generated. Then read the policy that governs your submission, because permission and disclosure are set there and nowhere else.

The institutional side is a different problem and this tool does not address it. The teaching centres and research guides in the result set converge on one recommendation, which is worth stating plainly even though it is not ours: detector output should not be the sole basis of a misconduct decision. Assignment design, explicit policy and direct conversation are what those sources recommend in its place.