How Teachers Check for AI Writing and Why Detectors Get It Wrong

A student writes an essay over 3 evenings, turns it in and gets an email: the AI detector says it is 40% AI-generated. The student did not use AI. Now what?

This is happening in classrooms across the country. AI detection tools are widely available, but they are far less certain than their percentage scores suggest. Teachers need a fair way to spot misuse, and students need a way to prove their work is their own.

This guide explains how teachers check for AI writing, what the research says about detector accuracy, and the simple habits that let any student show their process.

How teachers check for AI writing

Most teachers use a mix of methods rather than one tool. The common ones are:

1. AI detection software. Tools such as Turnitin’s AI writing indicator, GPTZero and others estimate how much of a text looks machine-generated.
2. Comparison with past work. A sudden jump in vocabulary, structure or polish compared to a student’s in-class writing raises questions.
3. Content checks. Made-up quotes, fake citations and confident but wrong facts are classic signs of AI output. Our explainer on AI hallucinations covers why chatbots invent sources.
4. Process evidence. Drafts, outlines, notes and document version history show how the work came together.
5. Conversation. Asking a student to walk through their argument, explain a word choice or summarize a source is often the most revealing check of all.

The best teachers treat a detector score as a reason to look closer, not as a verdict.

How AI detectors work

AI detectors are themselves AI models. They are trained to estimate how likely it is that a passage was produced by a large language model rather than a person.

Many rely on statistical patterns. Language models tend to choose predictable words, so detectors often measure:

• Perplexity: how predictable each word is. Lower perplexity looks more machine-like.
• Burstiness: how much sentence length and structure vary. Human writing tends to be more uneven.

The problem is that plenty of humans also write predictably. Students taught to follow a formula, writers using simple vocabulary, and people writing in a second language can all produce text that “looks” like AI to a statistical model.

What the research says about AI detector accuracy

The evidence on detector accuracy is the strongest reason for caution.

OpenAI shut down its own detector. OpenAI launched an AI text classifier in January 2023. In its own tests, the tool correctly identified only 26% of AI-written text as “likely AI-written” and wrongly labeled human writing as AI 9% of the time. OpenAI discontinued it on July 20, 2023, citing its low rate of accuracy.

Detectors are biased against non-native English writers. A 2023 Stanford study by Liang and colleagues, published in the journal Patterns, tested 7 widely used detectors. They misclassified more than half of the TOEFL essays written by non-native English speakers as AI-generated, with an average false positive rate of about 61%. The same detectors were nearly perfect on essays by US 8th graders. The researchers linked the bias to the lower perplexity of writing with a more limited vocabulary.

Vendors acknowledge limits at low scores. Turnitin says its document-level false positive rate is below 1% for documents with more than 20% AI writing, a vendor figure rather than an independent measurement. Because false positives are more common at lower levels, Turnitin’s reports now show an asterisk instead of a number for scores below 20%.

Some universities turned detection off. Vanderbilt University disabled Turnitin’s AI detector on August 16, 2023. It noted that even a 1% false positive rate, applied to the roughly 75,000 papers its students submitted in 2022, could have wrongly flagged around 750 papers.

Key takeaway: An AI detector score is a probability estimate, not evidence. It can flag honest students and miss edited AI text. Treat it as a prompt for a conversation, never as proof on its own.

Why false positives happen

False positives are not random glitches. They come from how detectors are built.

• Overlapping styles. Clear, grammatical, well-organized writing is what schools teach, and it is also what AI produces.
• Training gaps. Detectors are trained on certain kinds of text. Writing that differs from that training data, like essays from English learners or specialized technical writing, gets misread.
• Moving targets. New AI models write differently from older ones, so detectors are always catching up.
• Mixed authorship. When a student uses spell check, a grammar tool or an AI tool to brainstorm, the line between “human” and “AI” text blurs.

False negatives matter too. A student who lightly edits AI output can often slip past detection, which means detectors can punish honest students while missing dishonest ones.

What teachers can do instead of relying on detectors

Teachers who want a fair process usually combine clear policy with assignment design.

• Publish a clear AI policy for each assignment: allowed, allowed with disclosure, or not allowed.
• Grade the process. Collect a proposal, an outline and a draft before the final version.
• Use in-class writing to build a baseline of each student’s voice.
• Ask for reflection. A short note on how the student researched and revised the piece is hard to fake.
• Hold a conversation, not an accusation, when something looks off. Ask the student to explain their argument and sources.
• Document decisions and follow your school’s academic integrity process rather than acting on a score.

For teachers exploring AI in a positive way, our guides on AI in education and how teachers use AI to plan lessons show practical classroom uses.

How students can document their writing process

The best defense against a false accusation is evidence you create while you work. These steps take little extra time.

1. Write in a tool with version history. Google Docs (File, then Version history) and Microsoft Word with OneDrive autosave both record how a document changed over time. Do not write elsewhere and paste the final text in, because that erases the trail.
2. Keep your notes and outlines. Save research notes, brainstorm lists and outlines, even rough ones.
3. Save sources as you go. A bookmark folder or reference list with dates shows real research.
4. Use process-tracking tools if your school allows them. Grammarly Authorship, for example, labels text as typed, pasted or AI-generated and can replay how a document was written. Students control when it records.
5. Follow the AI policy and disclose use. If AI is allowed for brainstorming or feedback, say how you used it. Our guide on how students can use AI for homework explains honest ways to do that.
6. Be ready to talk about your work. If you can explain your thesis, your sources and why you made key choices, that is strong evidence of authorship.

What to do if you are falsely accused

If a detector flags your work and you did not use AI, stay calm and be organized.

1. Ask for specifics: which tool, what score, and which passages were flagged.
2. Share your version history, drafts, notes and sources.
3. Offer to discuss the work or write a short sample under supervision.
4. Point to the research on detector limits, including OpenAI’s withdrawn classifier and the Stanford study on non-native writers.
5. Follow the formal appeal process in your school’s academic integrity policy, and ask a parent, advisor or student advocate to help if needed.

Where to start

For teachers: write a one-paragraph AI policy for your next assignment, require a draft checkpoint, and decide in advance that a detector score alone will never decide an outcome.

For students: start your next assignment in a document with version history turned on, keep your notes in one folder, and read your class AI policy before you open a chatbot.

For schools: review whether your detection tool’s limits are explained to staff, and make sure the integrity process requires more than a software score. If your district is building an AI policy, the structure in our AI acceptable use policy guide can be adapted for classrooms.

Conclusion

AI detectors can help a teacher decide where to look, but they cannot tell anyone with certainty who wrote a piece of text. The published evidence shows real false positive risk, especially for English learners and short or formulaic writing. A fair approach pairs clear policies with process evidence and honest conversation. Students who keep their drafts and version history have the strongest protection of all.

Sources

• OpenAI: New AI classifier for indicating AI-written text — 26% true positive and 9% false positive rates, and the July 20, 2023 shutdown note
• Liang et al., GPT detectors are biased against non-native English writers (arXiv) — about 61% average false positive rate on TOEFL essays across 7 detectors
• Turnitin Guides: AI writing detection in the classic report view — asterisk display for scores below 20%
• Vanderbilt University: Guidance on AI detection and why we’re disabling Turnitin’s AI detector — August 16, 2023 decision and false positive reasoning
• Grammarly Authorship — how the process-tracking tool labels typed, pasted and AI text