Catching AI in Student Work: The Uncomfortable Truth

What this checklist can and cannot do

No method proves that a student used AI. Detection software, your own judgment and this checklist cannot prove it. A checklist gives you a consistent way to select the papers that need a closer look. It also prepares you for a fair, private conversation with the student. Use the checklist to collect evidence. Then let the conversation and the academic integrity process of your school decide.

Signals that only the teacher can see

You know the voices, habits and class work of your students. Software cannot do this part of the review, so these signals carry the most weight.

Item	Question to ask
Voice	Does the paper sound like the in-class writing and the earlier assignments of the student?
Knowledge level	Does the depth of the analysis match the work of the student in class?
Course material	Does the paper use the assigned readings and the specific angle of your prompt?
Class discussion	Does it use ideas from your lessons, or only general points that the class did not discuss?
Personal connection	If the prompt asked for opinions or experiences, does the paper include real ones?
Process evidence	Can the student show notes, an outline, drafts or a version history?

A change in voice does not prove AI use. Heavy editing by a tutor, help from a parent or copying from another student can cause the same change. A version history helps, but a student can create one by typing from a finished text. Use it as one piece of evidence.

Text patterns are weaker signals

Detection software looks for the patterns in this list. A careful human writer can show all of them. Use them only as supporting evidence:

• Smooth prose with no errors and no personal voice
• Citations that do not exist, or sources that do not say what the paper claims
• Paragraphs with almost the same length and structure
• General statements about the topic that ignore the specific prompt
• Many hedges, for example “one might argue” or “there are various perspectives”

What detection scores tell you

Detection tools calculate how “machine-like” a text is. They do not know how the student wrote it.

A 2023 study from Stanford (Liang et al.) tested seven widely used detectors on 91 essays from the TOEFL English exam. Non-native English speakers wrote these essays. On average, the detectors labeled 61% of the essays as AI-generated. The same detectors were almost perfectly accurate on essays by US eighth-grade students.

In June 2023, Turnitin reported a false positive rate of about 4% at the sentence level. A false positive is human text that the tool marks as AI. At the document level, Turnitin reported a rate below 1% for documents with 20% or more AI writing. In a large course, even 1% means that the tool flags some students who did their own work.

The educator FAQ from OpenAI says that its research did not find AI detectors reliable enough for decisions about students. It also says that ChatGPT has no knowledge of which content is AI-generated. If you ask ChatGPT “did you write this?”, it can make up an answer.

If your school provides a detection tool, record the score. Do not use the score as a verdict.

Step-by-step review of a suspicious paper

1. Compare the paper with the in-class writing or earlier work of the student.
2. Write down specific differences in vocabulary, sentence length and voice.
3. Go through the table of teacher signals above.
4. Look up each source that the paper cites.
5. If your school provides a detection tool, record its score as one data point.
6. Talk with the student in private.
7. Write down what you compared, what you found and what the student said.
8. Follow the academic integrity policy of your school for the next step.

In the private conversation, ask the student to do these tasks:

• Explain the thesis in their own words.
• Describe their research and writing process.
• Discuss a specific paragraph and the choices in it.
• Show their notes, outlines or drafts.

A student who wrote the paper can usually discuss it in detail. If a student cannot explain the thesis of their own paper, that is stronger evidence than any score.

What not to do

• Do not use a detection score as proof.
• Do not paste a paper into ChatGPT or another chatbot and ask if it wrote the paper.
• Do not accuse a student in front of the class or in a public message.
• Do not ignore the higher risk of false positives for English learners and for writers with a plain, formal style.
• Do not skip the conversation.

How to start the conversation

Start with questions, not accusations. For example, say: “I see some differences between this paper and your earlier work. Tell me how you wrote it.” A sentence such as “I think that you used ChatGPT” puts the student on the defensive. These questions also help:

• “Walk me through how you came up with your thesis.”
• “Which source helped you most, and why?”
• “What was the hardest part of writing this?”
• “You used the term [term]. What does it mean to you?”

Assignment design that makes the question easier

These assignment designs give you more evidence about the writing process of each student:

1. Ask students to connect their argument to a specific class discussion or reading.
2. Collect outlines, drafts or research notes at set dates before the final paper.
3. Collect a timed, in-class writing sample early in the term as a baseline.
4. Write narrow prompts about specific passages, not broad themes.
5. Add a short oral part where each student explains the paper.

Use this checklist with Claude or ChatGPT

Do not use Claude or ChatGPT as an AI detector. They cannot tell you if AI wrote a text. They can help with the steps of this checklist that collect evidence. Paste the checklist, the paper and, if you have one, an earlier writing sample into a new chat. Then ask for a structured comparison. For example:

Below is my review checklist, a student's essay and an in-class
writing sample by the same student. Do not judge whether AI was
used. Instead: 1) list concrete differences in vocabulary, sentence
length and style, with quotes from both texts; 2) list every cited
source so I can look up each one; 3) suggest five questions about
specific paragraphs that I can ask the student.

Before you act on the answer, read these cautions:

• Make sure that each finding is correct. Compare it with the texts yourself. The AI can miss differences, invent them, or describe a citation incorrectly.
• Remove student names and other identifying details before you upload student work. Follow the rules of your school for sharing student work with outside services.
• Do not let the tool decide on misconduct. That decision belongs to you and to the process of your school.

A docs portal with Ask AI
and support tickets

Connect a repository, or describe your product to start. Each docs site has search, Ask AI, an API console, an MCP server and support tickets.

• Publishes from GitHub on every push
• Works with Mintlify repositories
• API console from your OpenAPI file
• MCP server and llms.txt for AI agents
• Ask AI and support tickets
• Analytics for people and AI agents

Start Free→ Questions? Ask AI in our docs

No card required.