How Do Teachers Check for AI? A Practical Guide to Detection and Classroom Policy

Every teacher with a stack of essays has had the same moment: a paragraph that reads a little too evenly, a vocabulary jump that does not match the student in front of you, an argument with no fingerprints on it. The question that follows — how do I actually check this? — has become one of the most common questions in education right now, and the honest answer is messier than the marketing on any detection tool suggests.

This guide covers what the tools measure, where they fail, and what a workable classroom policy looks like when you accept that certainty is not available.

What an AI detector actually measures

AI detectors do not recognize text that a model wrote. They estimate how predictable the text is. Language models generate the statistically likely next word, so their output tends to be smoother and less surprising than human writing. Detectors score two things:

• Perplexity — how surprising the word choices are. Low perplexity (very predictable) leans machine.
• Burstiness — how much sentence length and rhythm vary. Human writing tends to lurch; machine writing tends to hum along evenly.

That is the whole trick. Which explains the failure mode: any writer who is careful, formulaic, or writing in a second language produces low-perplexity, low-burstiness text. So does a student following a five-paragraph essay template you taught them. The detector cannot tell disciplined prose from generated prose, because on the metric it measures, they look the same.

“A detector score is a hypothesis, not a verdict. Treat it the way you would treat a smoke alarm — worth walking into the room, not worth calling it a fire.”

Comparing the main AI detectors for teachers

The tools most schools land on differ less in accuracy than in workflow fit. Here is how they actually differ in daily use:

• Turnitin AI writing detection — bundled with plagiarism checking and integrated into most LMS gradebooks. Best fit if your institution already uses Turnitin; the AI score sits next to the similarity report so you are not juggling tools. Reports sentence-level highlighting and a percentage of the document flagged.
• GPTZero — the most common standalone choice for individual teachers. Free tier for short documents, paid tier for bulk and classroom management. Offers a writing-process replay feature when students draft inside its editor, which is more useful than the score itself.
• Copyleaks — strongest multilingual coverage, which matters if you teach English learners or world languages. Sells an LMS integration and an API, so it scales to district level.
• Originality.ai — built for publishers rather than classrooms, so it is aggressive by design. High sensitivity means more false positives on student work; better suited to checking submitted content than grading a section of ninth graders.
• Google Docs / Word version history — not a detector at all, and the single most useful signal available. Free, already in your workflow, and shows whether a document was written or pasted.

Independent testing has repeatedly found that detectors flag human-written text as AI at rates high enough to matter in a class of thirty — and that light paraphrasing collapses their accuracy. Vendors have quietly moved from accuracy claims to language about "indicators" for exactly this reason. Several universities have disabled institution-wide AI detection outright rather than defend individual scores.

The signals that beat detectors

If detection scores are weak evidence, what is strong? Process. AI-generated work has no history; student work does.

1. Version history. Open the Google Docs revision timeline. Real drafting looks like hundreds of small edits over time. Pasted work looks like one enormous insertion at 11:47 p.m.
2. Voice mismatch. Compare against a piece of in-class handwritten or timed writing from the same student. You are not looking for quality — you are looking for a different person.
3. Fabricated specifics. Ask about a cited source, a quoted page, or a data point. Invented citations remain one of the most common tells.
4. Missing course context. Generated work rarely references the class discussion from Tuesday, the specific edition you assigned, or the framework you spent a week on.
5. The conversation. Ask the student to walk you through one paragraph: why this claim here, what they cut, what they almost argued instead. Students who wrote it can do this in ninety seconds.

A pedagogical framework for managing AI-generated work

Detection is downstream of design. The classrooms with the fewest integrity fights are not the ones with the best detector — they are the ones where the assignment makes AI-only submission an obviously bad strategy. Four moves do most of the work:

1. Make the policy specific, per assignment

"No AI" is unenforceable and students read it as noise. Label each assignment with a tier: AI-free (done in class or on paper), AI-assisted (allowed for brainstorming, outlining, or feedback, with disclosure), or AI-collaborative (expected, with the transcript submitted). Students follow rules they can actually apply.

2. Grade the process, not just the artifact

Require an outline, an annotated draft, a source log, or a two-minute recorded walkthrough. Weight those at 30–40% of the grade. This does not detect AI; it removes the incentive, because the shortcut no longer saves time.

3. Anchor tasks in things a model cannot know

Local data, a lab result the class generated, a text from Thursday's discussion, a personal interview, an in-class debate the student must respond to. Specificity is a better defense than surveillance.

4. Teach the tool before you police it

Run one session where the class prompts a model, then critiques the output together — the vague thesis, the confident wrong citation, the flattened voice. Students who have seen a model fail in front of them trust it less, and the conversation shifts from rule-breaking to judgment.

When you suspect AI: a defensible sequence

1. Pause before accusing. Write nothing on the paper and enter nothing in the gradebook yet.
2. Gather process evidence — version history, prior samples, source checks. Detector output is one input among several, never the basis for a finding.
3. Open with curiosity, not a charge: "Walk me through how you approached this." Ask about choices, not intent.
4. Offer a resubmission path when the evidence is ambiguous — which it usually is. An oral defense or supervised rewrite resolves most cases without a formal process.
5. Escalate only with process evidence, and document what you observed rather than what a tool scored.

The false-positive cost is asymmetric. A missed case of AI use costs one assignment's worth of learning. A wrong accusation costs a student's standing and your relationship with the class. Set your threshold accordingly.

The short version

• Detectors estimate predictability, not authorship — the score is a hypothesis.
• Version history is free, already in your workflow, and more informative than any detector.
• Design assignments so the shortcut stops saving time; that beats surveillance.
• Talk to the student before you conclude anything. Most ambiguity resolves in one conversation.