AI Adoption GuideEducationAssess
AI-content Detection
A classifier combines AI-generated content detection with writing pattern analysis to flag submissions for academic integrity review.
Education processRecruitAdmitEnrollTeachAssessCredentialGraduateAdvance
By Don, DoneThat’s AI coach · updated
A flag is a review packet, not a misconduct finding
The classifier should return a flag with cites to the detector output and to the writing-pattern signals used. That packet is a reason to open academic-integrity review. It is not a finding of misconduct.
Integrity officers and designated faculty still own the case. They decide whether to request drafts, hold a process conversation, or close with no action. Auto-failing a student on classifier output skips that ownership and writes an uncertain instrument into the grade before anyone has checked the assignment, the file, and local policy.
Writing-pattern analysis belongs in the same packet, not in a second verdict. A sudden register shift, flattened cadence, or a stretch that does not match the rest of the essay can be listed as a signal with a cite. Those signals still need a person who knows the prompt and what the campus treats as unauthorized aid. Time pressure, writing-center edits, translation, and dictated drafts can move the same surface features.
Do not invent a percent-AI number the detector did not emit. If the tool returns a qualitative label, a binary flag, or highlighted spans, quote that output. A homemade percentage makes the packet look more precise than the instrument and is hard to unwind in a hearing.
Run the classifier on the submitted file
Run the check on the file in assessment intake, not on a paste from email or a screenshot of a draft. Typical homes for that file are learning platforms in the Canvas, Blackboard, Moodle, and Anthology class, and integrity tools in the Turnitin class. Treat those products as the systems that already store submissions and related reports. This page does not rank them or ascribe features to any one of them.
Repeatable run:
- Confirm the assignment is in scope: individual written prose, syllabus notice, and local policy that allows this class of check.
- Isolate the student text. Strip instructor comments and other students' work.
- Run the classifier once on that file.
- Store the raw detector output exactly as emitted: labels, spans, identifiers, timestamps if present.
- Store the writing-pattern signals with locations a reviewer can find.
- If the result is a flag, attach both cite streams and route to human review. If there is no flag, do not open an integrity case from this tool alone.
Do not fold detection into rubric-based essay scoring. The rubric answers whether the essay met the assignment. The classifier answers whether the text should be reviewed for integrity. Combining them in one pass trains staff to treat style as a scoring defect.
Keep detection out of coaching threads. Formative feedback generation is for revision. A flag in the same comment set reads as an accusation offered as advice.
Attach detector output and writing-pattern cites
The packet is usable only if both cite streams are attached and neither is rewritten into stronger language.
Detector cites point at what the AI-generated-content detector returned. Copy the labels and highlights. "Flagged" is not "generated by a chatbot." "Paragraphs 4 and 5 highlighted" is not "the student used an AI assistant."
Writing-pattern cites point at which features fired and on which passages. Keep comparisons inside this submission unless your process already has a consented baseline of that student's prior coursework. Do not scrape other courses to build a shadow profile.
Illustrative example: a take-home midterm arrives as a long essay. The detector marks two consecutive body paragraphs and returns a flag, not a percentage. The writing-pattern layer cites a drop in first-person hedging in those same paragraphs and a vocabulary band that does not appear in the opening or closing sections. The packet lists those spans and the raw labels. The officer reads the prompt, sees that the flagged stretch was supposed to be personal reflection, and schedules a process conversation. Nothing in the packet fails the student. Drafts, the meeting, and policy decide the next step.
If the campus already uses reviewer decision assistance for integrity or appeals, load this packet as cited evidence items, not as a recommended sanction.
Leave thin files unscored
Empty stays empty if the file is too short to score. Discussion replies, formula-only uploads, caption-length reflections, and stub documents do not become "0% AI" and they do not become "clear." They become no score.
A blank result is not a clean bill of health. It is a statement that the instrument did not have enough text. Do not treat silence as innocence, and do not re-run the same stub until a number appears. Re-running a thin file until the tool speaks is how invented scores enter the record.
If policy still requires an integrity check on short work, use a human sample, an oral check, or a process artifact such as notes or a lab book. Do not force a classifier output.
Record the skip: why the file was ineligible, and that no flag and no clearance were issued. That note blocks a later claim that detection ran and found nothing when nothing valid ran.
Hand every flag to a human reviewer
Name the owner before the term starts: the instructor of record, a departmental integrity lead, or the campus academic-integrity office. The classifier does not close the case.
The reviewer should receive the assignment prompt, the full submission, the detector output, the writing-pattern cites, and any process materials policy allows. They should not receive a dashboard state that already reads as misconduct.
The review is a process check, not a second pass of the detector. Ask how the essay was produced, compare any drafts you already have, and weigh the cited spans against the prompt. Close the case when the packet does not support a charge. Refer only when policy and the conversation warrant it.
Auto-failing a student from the flag is the failure this workflow exists to prevent. Grade retention, resubmission, oral defense, and disciplinary referral are human decisions. Do not let automation write a failing grade or an honesty violation into the gradebook from classifier output.
Keep the student-facing notice modest: the submission was flagged for review, how to share drafts or meet, and where policy lives. Do not paste detector jargon or a homemade AI percentage into that notice.
Keep course detection separate from admissions screening
This classifier flags a course assessment for integrity review. Essay authenticity screening for admissions is a different workflow, with different consent, different baselines, and different harm if the signal is wrong.
Do not reuse a course flag, a writing-pattern profile, or a detector span from this pipeline as evidence in an admissions, scholarship, or placement file. Do not run the course classifier on application essays because the model is already on campus. Mixing the two turns a teaching-term integrity signal into a gateway decision the applicant did not agree to, and it contaminates both records.
If a unit asks for "the AI score" from a course to inform admission, refuse the number, especially a number the detector never emitted, and point them at their own authenticity process.
Is this worth automating for you?
Whether this pays back depends on how much time it takes your team today. Most teams estimate that from memory, and the estimate is usually wrong in one direction or the other.
DoneThat reconstructs where the time actually went, with no timers to forget, so you can measure the baseline before committing to a project and check the gain afterward.
Measure the baseline first