Contract cheating signals: the essay mill leaves fingerprints a similarity score cannot see
Document history is the quiet witness
Real drafts accrete. They show edits over days, pasted fragments with formatting scars, dead ends deleted, and a version trail in the learning platform or the document editor. A purchased essay often arrives as a single clean file, created in one session, sometimes carrying an author name in its properties that is not the student, or a template and language settings from another country. None of that proves purchase alone, and all of it justifies the next step, which is a conversation, not a verdict.
Set the expectation before the assessment, in the course rules: drafts and working notes may be requested, and document history is part of the submission record for take home work. Students who write their own work keep that history without thinking about it. A rule announced in advance is fair process. The same request invented after suspicion is an argument about targeting waiting to happen.
Style breaks are a lead, not a conviction
A sudden change in register, citation habits, or competence between a student's timed writing and their take home essay is a signal worth one follow up, nothing more. English language learners progress unevenly, tutoring legitimately improves structure, and a student may simply care more about one topic. Institutions that treat a style model score as evidence have built an accusation machine with a false positive rate they have not measured and cannot defend in an appeal hearing.
The defensible use is triage: the signal selects work for a human process, a short viva or a supervised rewrite of a passage, applied under a published policy and ideally to a sample of all submissions, not only the flagged ones. A process only ever applied to the accused is not a process. It is a conclusion with paperwork.
The five minute conversation does the real work
Ask the student to explain their thesis, why they chose their main source, and what they cut. An author answers from memory, including the wrong turns. A purchaser answers from the text, reading back the introduction with growing vagueness about everything not in it. Five minutes of specific questions separates the two more reliably than any software score, and it gives the honest student a fair, human chance to be believed.
Record the conversation's questions and answers factually, and decide on the whole record: history, style, conversation, prior work. Panels that see the full ladder make decisions that survive appeal. Panels handed a single percentage make the news.
Assessment design starves the mills: staged submissions with a marked proposal and draft, personalised data sets or local cases a mill cannot stock, a short in class writing baseline for every student, and questions that require the course's own materials rather than a generic essay plan. Mills sell scale and speed. An assignment built from this term's seminar discussion is a bad product for them to fulfil, and they will still sell it, badly, which is the point.
Prevention beats detection, again
Say plainly why the design exists. Students mostly hate contract cheating, because it devalues their own work. A course that explains its design honestly gets cooperation a surveillance pitch never earns.
Frequently asked questions
Why does similarity software miss contract cheating?
Because the submitted text was written for that assignment and has not been published or submitted before, so there is nothing to match against. Similarity measures copying. Contract cheating is an authorship problem and needs authorship evidence.
Why does similarity software miss contract cheating?
No. Metadata can justify an investigation and support a case built from several signals, but students share devices, templates, and help legitimately. A finding needs the combined record, usually including the student's own account in a fair process.
Why does similarity software miss contract cheating?
No. Their error rates, especially against non native English writing, make them unsafe as evidence. Use authorship signals to select work for a human, documented process, never as the process itself.
Why does similarity software miss contract cheating?
A short structured conversation about the work's choices and sources, under a policy announced in advance and applied consistently. Authors pass it easily and purchasers rarely do, which is the combination a fair check needs.