A submission looks off — too clean, too fast, too unlike anything this student has turned in before. What you do next matters twice: for the grade, and for the student. This page lays out a review workflow you can follow and defend: what to check by hand, when a free screening scan adds a useful signal, and how to raise it with the student without turning a suspicion into an accusation.
One expectation up front: no workflow — manual or automated — settles authorship for certain. The goal is a fair process that surfaces the files worth a closer look, not a machine verdict.
You hold the strongest signal before any tool runs: this student's history. Pull up their earlier submissions and compare — naming habits, comment density, the mistakes they usually make, the way they structure a solution. A sudden jump in polish, or an idiom you have never seen from them, is worth noting. A stranger cannot run this check; you can.
Write down what struck you before scanning anything. That baseline is the context every later signal gets read against — including the detector's.
Before you reach for a tool, read the file yourself. Four signals are worth checking — each one is a reason to look closer, never a conclusion.
For the full walkthrough of each signal — including the ways honest code trips them — read How to tell if code was written by AI.
Read the four signals in depth →
Intro courses lean on a few languages — see what AI-generated Python code looks like, and the Java patterns that cut both ways.
When manual reading leaves you unsure, add one independent signal. The point is not to outsource the judgment — it is to check whether the parts of the file you flagged by hand are the same parts a detector flags.
Grading a whole class set? Batch mode screens all the files from one assignment and ranks them for review — see AI code screening for teachers.
Want to see the output first? Here is a sample report from a full assignment set.
Screening signal only. Review flagged code with the student before you act on it.
The score estimates how closely the code matches patterns associated with AI generation — it is a screening signal, not evidence of misconduct. In a grading context, three situations raise false-positive risk: very short assignments (files under 10 non-empty lines are not scored at all), heavily templated starter code the whole class shares, and courses with strict style rules that push every student's code toward the same uniformity. A flag on any of these calls for extra skepticism, not less process.
If the signals stack up, talk to the student before anything else. Keep it concrete and non-accusatory: ask them to walk you through the flagged snippets — why this approach, what this function returns, how they would extend it if the assignment changed. A student who wrote the code can usually explain it; a student who hesitates may still have a fair reason, so treat the conversation as one more signal, not a confession.
Document what you reviewed and what was said. Whatever your institution's integrity process looks like, a detector score belongs at the start of it — never at the center.
No. A score is a screening signal that helps you decide where to look — it is not evidence of AI use. Build any case on your own review of the code and your conversation with the student, and follow your institution's process for what happens after.
Yes. Instructors who use screening tools usually say so in the syllabus and explain how flagged work is reviewed — including the student's chance to explain their code. Transparency up front reduces both anxiety and disputes later.
Usually it says more about the assignment than the class. Shared starter code, strict templates, and very constrained problems make honest submissions look alike — and look like model output. Check whether the flagged snippets are the shared template before reading anything into the scores.
Scanning the whole set is fairer and more useful. Suspicion-only scanning bakes in your biases and gives you nothing to compare against. A full-class scan shows which files genuinely stand out from the class baseline.
Run a free single-file check now, then review it with the signals above. When the next assignment lands, batch mode screens the whole set and ranks it for review.