ChatGPT is in every classroom. Here's a practical, researcher-informed workflow for teachers — how to detect AI-generated work, how to talk to students about it, and how to build policy that works without false-accusation risk.
By 2025 the majority of students have used an LLM for some part of their academic writing. Surveys of university students consistently put that number between 60% and 90% depending on discipline and country. The question is no longer whether students use AI but how much, for which tasks, and with what consequences.
The academic-integrity question splits into two sub-questions. Is a given submission AI-generated? — a detection problem. Does AI use violate the assignment's rules? — a policy problem. Teachers need answers to both, and the order matters: policy comes first, detection confirms.
Running detection without clear policy creates false-accusation risk. Running policy without detection creates honour-system cheating. The practical answer is a joint workflow where both layers support each other.
Good AI policy is explicit on four dimensions. What's allowed: brainstorming, outlining, grammar checking, reference hunting — commonly permitted even by strict policies. What's prohibited: whole-sentence or whole-paragraph generation submitted as the student's own work. What must be disclosed: any AI-assisted task, logged in a disclosure statement at submission. What's the consequence: academic-integrity tribunal, grade penalty, resubmission, or escalation — state it up front.
Publish the policy before any AI-detector scan is run on a submission. Students who are told “we will detect AI” after submission have a legitimate grievance; students who are told “here is the policy, and here is how we verify” at the start of term cannot. Treat detection as enforcement of a published policy, not a surprise.
Align with your institution. If your university has a model policy, adopt it. If it doesn't, borrow from the MLA, IEEE, or your national regulator. Inconsistency across teachers in the same institution creates student grievance and legal risk — align faculty before rolling out detection.
An AI-detection score is a signal, not a verdict. A 92% AI probability on a submission is a strong reason to investigate further — it is not proof. Our accuracy benchmark is honest about this: at the 50% threshold we aim for 0 false positives on our validation set, but your students' writing is not our validation set.
Combine the score with three other signals before any decision. Writing history: does this match the student's previous submissions? In-class signals: in-class essays, oral discussion, short-answer quizzes — do they match the submission's level? Technical context: submission timestamp, edit history (if the platform exposes it), any unusual metadata.
A score plus at least one corroborating signal is an investigation-worthy case. A score alone is a flag, not a finding. This rule — originally documented in academic-integrity literature long before AI — protects both students and teachers and is the single most effective lever against false-accusation disputes.
If a submission scores as likely AI, meet with the student. Do not open with the accusation. Open with the work. Ask the student to walk you through their process: what they researched, what their draft looked like, what they changed. Students who wrote the work can answer these questions fluently. Students who used AI often cannot — not because they are dishonest, but because they have not engaged with the material.
The purpose of this conversation is to gather evidence, not to trap. Take notes on what the student says. If the conversation resolves the flag — their process is coherent and their draft history matches — the flag is retracted. If the conversation reveals inconsistencies, you now have corroborating evidence to proceed formally.
Avoid these common errors. Do not lead with the detector score — students will feel ambushed. Do not treat the score as confession-worthy — some students will admit under pressure even when innocent. Do document every conversation — your institution's due-process requires written records.
AI detection finds generated text. Plagiarism detection finds copied text. Students submit a mix of both — some LLM-drafted paragraphs, some copy-pasted from other sources, some genuinely original writing. A workflow that only scans for AI misses copy-paste; a workflow that only scans for plagiarism misses fully generated content.
Our desktop Plagiarism Detector runs both in a single scan: one pass for matches against 4 billion indexed web pages, academic databases, and the institutional PDAS corpus, plus the same AI-detection engine that powers our online tool. Combined verdict per document in under a minute.
For institutions that prefer browser-based workflows, our free online tool covers AI detection and the free-demo desktop download adds the full source-matching passes. Most universities run some mixture of the two depending on faculty workflow.
Paste a sample submission and see the per-sentence verdict. Classroom-ready. No signup, no cloud storage.
Disclosure-first: any AI use requires a short statement at submission — “I used GPT-4 to outline section 2 and edit section 3 for grammar.” No detection penalty if disclosed; full penalty if undisclosed AI is detected. Low-friction for students, high-accountability.
AI-free assignments: clearly marked submissions that must be written entirely without AI. In-class, oral, or proctored. Used for final exams, diagnostic writing, and any task where AI is beside the learning objective.
AI-permitted assignments: explicitly allow AI as a research or editing tool; grade the student's final work on quality regardless of how it was produced. Students learn to use the tool; teachers grade the outcome. This approach has the highest faculty adoption and lowest detection workload.
You will miss some AI-generated submissions. Humaniser tools, short assignments, and hybrid human-AI writing all defeat text-based detection at current generator levels. Accept that the goal is not 100% detection but meaningful deterrence and fair handling of flagged cases.
You will flag some human submissions as AI. Non-native English writing, heavily edited academic prose, and some genuinely unusual student styles all score higher than expected. The 0-false-positive number in our benchmark is on the validation set; your students are not that set. Combine with corroborating signals before any action.
The workflow that works sustainably: publish policy, run detection at submission, flag high scores for investigation, investigate with the student present, document everything, escalate only when corroborated. Teachers who follow this sequence report both reduced AI usage and reduced false-accusation disputes within one semester.
This article is educational guidance, not legal advice. Academic-integrity policies and the legality of automated detection vary by jurisdiction and institution. Consult your institution's data-protection officer before deploying any detection workflow.
Stand with Ukraine!