Claude assignment checker

The draft you paste is sent to our server and analysed there. What comes back is prose: observable qualities of the writing, stated as findings, with the limits of the reading attached. The brief the server works from instructs it to describe observable signals and limitations and never to claim certainty about authorship, plagiarism or which model produced a text. That instruction is not a disclaimer bolted onto a score. It is the shape of the output.

Operationally: 5 runs per rolling 24 hours, an input bounded at 1 to 12,000 characters, an optional instruction capped at 300 characters, no account, no stored history. The button reads Analyze and there is exactly one result pane.

The instruction field is the variable most people leave empty and the one that changes the output most.

Step 1

Select a bounded unit of text: a section, an argument, an introduction. Unbounded input produces unbounded findings, which are harder to act on.

Step 2

State the question in the instruction field. A specific question constrains the analysis; an empty field leaves the engine to choose what is interesting.

Step 3

Treat the findings as a hypothesis about your draft and test each one against the text. Some will be wrong, and knowing which is the skill.

Grammarly reports the share of a document that "appears to be written with AI", split between AI-generated, "Resembles AI text" and "No AI text patterns found". QuillBot returns 0 to 100% with sentence highlights and explainer cards. Detector-checker.ai returns 0 to 100 in three bands. All three publish a version of the same accuracy claim, around 99% against the RAID benchmark.

Read the disclaimers on the same pages and the number stops meaning what it appears to mean. Grammarly writes that "currently there is no AI detector that can conclusively or definitively determine whether AI was used". Detector-checker.ai notes its published benchmark contains "no documented Claude or Gemini outputs", so a Claude-specific accuracy rate does not exist even in principle. A figure that cannot be interpreted is worse than no figure, because it will be acted on anyway.

The constraint is in the request itself. The analysis brief on this route forbids certainty about authorship, about plagiarism and about model attribution, and asks for findings plus limitations. An engine that cannot assert authorship cannot be quoted as having asserted it, which removes the failure mode that matters most in an academic setting.

The competing pages arrive at the same conclusion by a longer route. Detector-checker.ai states that its tool "does not identify the source model", cannot confirm Claude produced a passage and cannot distinguish Sonnet, Opus and Haiku from each other. It then ships a three-band verdict anyway. We would rather not produce the artifact that gets screenshotted into a misconduct email.

Scope is the honest part of any analysis tool. This one sees a string of text, so the following are structurally out of reach rather than merely unimplemented.

Brief compliance

The assignment question is not an input. Whether the draft answers it cannot be inferred from the draft alone.

Rubric performance

No criteria, no weighting, no predicted mark. Grading is a judgement against a document the tool never receives.

Source integrity

No reference index and no retrieval. A fabricated citation is indistinguishable from a real one at the level of text.

Similarity

No corpus, no matching, no similarity report. This is categorically not a plagiarism checker.

QuillBot documents its own weakness with unusual candour: "heavily paraphrased writing and highly formulaic content, such as legal jargon or academic writing, is more likely to present a false positive." Detector-checker.ai adds that "polished answers, cautious tone, consistent sentence structure" appear in careful human writing as readily as in machine output, and that editing moves results in either direction.

Put together, the properties that raise a detector score are the properties a marking rubric rewards. A student who has learned to hedge, structure and paraphrase well has learned to look guilty to an instrument that measures surface regularity. That is a reason to stop treating the instrument as evidence, not a reason to write worse.

No, and nothing in the result set claims otherwise once you read past the headline. Authorship is a fact about a process, and every tool here observes only the product. Text carries no signature, editing changes the surface, and two writers trained on the same conventions converge.

What does establish process is process evidence: drafts, version history, notes, a conversation. Institutional policy, as Grammarly correctly notes, is what governs. A checker belongs on the drafting side of that line, improving work before submission, which is the only use of it that is not contested.

No percentage, band, highlight or grade. No file upload, where Grammarly accepts a document and detector-checker.ai accepts 5 MB. No plagiarism database, no citation verification, no rubric input. No bulk queue, no API, no integration, no account, no saved history.

No accuracy figure either, because none has been measured here, and repeating someone else's benchmark would be borrowing a number for a different instrument. And no prediction of what Turnitin, GPTZero, Originality.ai or Copyleaks will report, with which there is no affiliation. The absent features are absent by decision, and the page lists them because a checker that hides its scope is the least trustworthy thing in this category.