# How to tell if a student essay was written by AI on Mac without false accusations

> Detector scores are not proof. To tell if a student essay was written by AI, check drafts, version history and a short talk with the student.

URL: https://www.lazur.app/for/teachers/problems/tell-student-essay-written-by-ai
Updated: 2026-11-09

Start with this: no AI detector gives you proof. A high score is a reason to look closer, not a verdict. This page shows how to check a student essay on your Mac, what each check can and cannot show, and what to do before you accuse anyone.

## Quick answer

- Detector scores are weak evidence, and they flag careful writing by non-native English writers.
- Check the drafts first. Pages and Google Docs keep version history that shows how the essay grew.
- Ask the student to explain one paragraph in their own words, and write down what they say.
- Set a written AI rule for each assignment, and let students disclose AI use without penalty.

## What judging whether a student essay is AI-written costs you

- **False accusations land on real students.** One in five students told the CDT survey that they, or someone they know, had been accused of relying on AI without any proof. Each accusation can mean a meeting, a lost grade or a record a student has to clear. The student bears that cost first, and it is hard to undo once a grade is in the system.
- **Non-native writers carry more of the risk.** In the 2023 study, 61.22% of human-written essays by non-native writers were flagged as AI. MIT Sloan cites about 5% of essays by native English writers flagged. If your class includes multilingual students, a detector score can single out students who write carefully in a second language.
- **Your evenings absorb the extra work.** No study I could find measures how long these checks take. Expect the time to go into reading histories, meeting students and writing notes. BLS says high school teachers often spend evenings and weekends grading papers and preparing lessons, so an investigation adds to that load.
- **Trust is harder to rebuild than a grade.** No study I could find measures trust directly. The practical risk is simple. A student who feels accused without a conversation may stop showing drafts or stop asking questions. Once that happens, the checks you rely on stop working.
- **The policy gap makes every case harder.** Gallup's 2026 survey found 18% of teachers get formal guidance on AI use from school administrators. 48% get only informal guidance, and 34% get none. Walton's summary of the 2025 Gallup survey says only 19% of teachers report a school AI policy. Without a written rule, each case becomes an argument about what the rule meant.

## Why it happens

1. **Detectors read style, not authorship.** A detector does not see who typed the words. It scores the text in front of it, and smooth, predictable prose scores higher. That makes style the main input. Clean, formal, evenly paced writing can score high whoever wrote it. The score answers a narrow question about text, not the question you care about, which is who did the work.
2. **Careful second-language writing looks predictable.** The 2023 study found that detectors consistently misclassify non-native English writing as AI-generated, while native writing is identified accurately. Careful, formal English often uses common phrasing. Common phrasing can read as predictable, and that is one way honest students get flagged.
3. **Detectors are easy to evade.** MIT Sloan says detectors are straightforward to evade by changing the prompt. A student who wants to hide AI use can often do it, while an honest student can still be flagged. The tool fails in both directions. That is why a detector is weak at proving either innocence or guilt.
4. **Take-home essays leave only a final file.** A take-home essay with only a final file leaves you no record of the work. The assignment asks for the product, so the product is all you can check. Adding drafts, checkpoints and a short talk gives you a record, and a reason to take the process seriously.
5. **Scores have become routine.** The CDT survey found 39% of teachers regularly use AI content detection software. When a tool is routine, its output can start to feel like a finding. A routine score gets treated as proof, even though MIT Sloan and the 2023 study both warn about error rates.
6. **Unclear rules about help.** Students and teachers often mean different things by AI help. Brainstorming, grammar fixes and translation can each count as help or as cheating, depending on the class. Without a written line, a student can break a rule they never knew existed. Your check then becomes a debate about the rule.

## How to fix it

- **Treat a score as a question.** Write down the score and the tool name, then stop. Do not share the result with the class, the parents or the student as a finding. Use it to decide where to look next. If your school provides a detector, ask what its error rates are before you rely on it.
- **Set checkpoints before the due date.** Ask for an outline, a rough draft and the final file as separate submissions. In Google Docs, share each file with edit access so the history is visible to you. Checkpoints give you a baseline for each student's writing, and they make the final essay less of a surprise.
- **Write one page of AI rules per assignment.** MIT Sloan recommends clear written AI-use policies that define cheating for generative AI. Give concrete examples of acceptable and unacceptable use for this assignment. A rule like brainstorm with AI, write every sentence yourself is easier to enforce than a vague instruction to use AI responsibly.
- **Let students disclose without penalty.** MIT Sloan suggests letting students disclose AI use without penalty. Disclosure is more likely when the disclosure itself is safe. Grade the essay on quality and process, and keep any disclosure separate from an integrity question. Students who disclose tend to give you cleaner evidence later.
- **Mix assessments on purpose.** Pair the take-home essay with in-class writing, short quizzes or a discussion. MIT Sloan warns that handwritten exams or oral presentations alone can raise equity concerns. A mix avoids relying on one format, and it gives students more than one way to show what they learned.
- **Compare with earlier in-class writing.** Keep a short writing sample from early in the term, such as a paragraph written in class. Compare tone, typical errors and vocabulary with the take-home essay. Expect change, since students grow over a term. Look for a sharp break from their usual work, not ordinary improvement.
- **Ask for Authorship reports only when shared.** If a student uses Grammarly Authorship, ask for the report link. By default, Authorship does not share a report. A link exists only when someone generates it and clicks Share. Report data stays on Grammarly's servers for up to 10 years unless deleted, so ask your school what it allows.
- **Give a clear response window.** Tell the student you have a question about the essay, and give them a set time to answer in writing or in person. Note the date you asked and the date they replied. A clear timeline protects you and the student if the case goes further.

You can tell if a student essay was written by AI with some confidence, but only by combining signals. Writing style on its own is weak evidence. Drafts, version history and a short conversation with the student carry far more weight. Start with the drafts, then add the rest.

![Teacher reading a student essay printout at a desk](https://images.lazur.app/content-images/problem/tell-student-essay-written-by-ai/tell-if-a-student-essay-was-written-1-78e5b4.jpg)

## Why a detector score is not evidence

AI detectors guess from patterns. They look for text that is smooth, even and predictable. Human writing can look that way too. A careful, formal essay can score high. So can a paragraph a student rewrote several times to sound polished.

The research here is uncomfortable. In a 2023 study, Liang and colleagues tested seven GPT detectors on 91 TOEFL essays written by non-native English writers. On average, 61.22% of those human-written essays were classified as AI-generated. Eighty-nine of the 91 were flagged by at least one detector. The same detectors identified native writing accurately, according to the authors.

- **61.22%** Average share of 91 human-written TOEFL essays by non-native writers that seven GPT detectors flagged as AI-generated ([source](https://arxiv.org/html/2304.02819v3))
- **39%** Teachers who said they regularly use AI content detection software, 2023-24 school year ([source](https://www.k12dive.com/news/student-teacher-ai-use-schools-cdt/737335/))
- **1 in 5** Students who said they or someone they know was accused of relying on AI to cheat without any proof ([source](https://www.k12dive.com/news/student-teacher-ai-use-schools-cdt/737335/))
- **18%** Teachers who report any formal guidance on AI use from school administrators, Gallup 2026 survey ([source](https://news.gallup.com/poll/710534/teachers-receive-no-formal-guidance.aspx))

> **A score is not a finding:** A detector can flag honest students and miss edited AI text. Never accuse a student on a score alone, and never share a score as a verdict.

MIT Sloan Teaching & Learning Technologies puts it plainly. Detection software has high error rates, and those errors can lead instructors to falsely accuse students. The same page says detectors are straightforward to evade by changing the prompt. So the tool misses some AI writing and flags some honest writing. A score cannot tell you which case you have.

## Check drafts with version history on a Mac

The most useful check is how the essay was built. A student who worked on it over several evenings leaves a trail of drafts, edits and half-finished sentences. A student who pasted a finished essay in one go leaves a different trail. Version history shows you that trail. It proves nothing alone, but it gives you something specific to look at.

If the essay is in Pages, open it and choose File > Revert To > Browse All Versions. Click Restore to make an earlier version current. Option-click Restore a Copy to open a copy in a new window. Work on the copy. Do not restore over the student's file, because that replaces their current version.

Google Docs works differently. Open the file and click Last edit at the top right of the page. Google shows the version history from there. You need permission to edit the file to see it. Viewers and commenters cannot see the history at all. For the full Pages routine, see [Pages for teachers on Mac](https://www.lazur.app/for/teachers/pages).

## A worked example of reading a history

Say a long essay arrives late at night. The history shows little activity for several days. Then one large block of text appears in a few minutes, with few typos and no deleted sentences. That pattern is worth a question. It is not a finding. Students paste from their own notes, from other files and from research they typed elsewhere.

Now look at the earlier drafts. Do they have the same ideas as the final version? Do they have clumsy sentences, corrections and paragraphs that were cut? Real drafts usually look messy in the middle. A file that goes from blank to polished in one step is not proof of AI. It is a reason to ask about the process.

Write down what you saw, with the dates and times from the history, before you talk to the student. Notes written after a conversation tend to reflect what you heard, not what you saw. Keep the notes in the same folder as the grading record.

![Student typing a student essay on a laptop in a classroom](https://images.lazur.app/content-images/problem/tell-student-essay-written-by-ai/tell-if-a-student-essay-was-written-2-89dca1.jpg)

## Edge cases that break the easy checks

- Writing in another app first. A student may draft in Word, then paste into Google Docs. The history shows one large paste, but the work is real.
- Dictation and accommodations. Speech to text can produce clean text with little editing. Check the accommodations on file before you read the history.
- Help from a tutor or family member. Outside help may or may not be allowed. Your policy should say which help counts.
- Grammarly Authorship. In Google Docs it needs the Grammarly extension in Chrome or Edge. In Word it works through the desktop app for Windows and Mac. By default it does not track writing activity, and a report is shared only when someone generates a link and clicks Share. The support page does not mention Safari.
- Non-native English writers. Formal, careful English can look predictable to a detector. Treat a high score on that writing as an even weaker signal, as the 2023 study suggests.

## Ask the student a few questions before you decide

Keep the talk short and calm. Ask the student to explain the thesis in their own words. Ask where one source came from and what it said. Ask which sentence they would change now. A student who wrote the essay can usually answer these quickly. Vague answers, or answers that do not match the text, are worth noting.

This check has limits too. A nervous student may freeze in a conversation, and that says nothing about who wrote the essay. MIT Sloan notes that oral presentations add stress for students with anxiety and for non-native English speakers. Offer a written follow-up before you draw a conclusion.

![Teacher talking with a student across a classroom table](https://images.lazur.app/content-images/problem/tell-student-essay-written-by-ai/tell-if-a-student-essay-was-written-3-338c3f.jpg)

## Ask for a process statement with each essay

A process statement is a short note the student attaches to the essay. It lists the tools used, what they asked, what the tool produced and what they changed. Keep it to a few sentences. Students who were honest can fill it in quickly. A vague statement is not proof of anything, but it tells you where to ask the next question.

Make the statement due with the final draft, so it arrives with the evidence you need. Add it to the assignment sheet, not to a separate email. If you grade it, grade it as part of the work. Students then treat it as a real part of the assignment.

| Signal | What it can show | How much weight it deserves |
| --- | --- | --- |
| Detector score | Predictable text patterns | Low. Flags careful second-language writing too. |
| Version history | How the text grew over time | Medium. A paste can come from real work. |
| Earlier writing sample | Voice, vocabulary and usual errors | Medium. Compare like with like, same task type. |
| Conversation | Whether the student can explain the work | Medium to high, when kept calm and fair. |
| Process statement | Which tools the student says they used | Useful when honest. A vague answer is a prompt to ask more. |

_How much weight each signal deserves. This ranking is my judgment from the sources, not a published rule._

## Mistakes teachers make with AI checks

- Treating a detector percentage as the chance the student cheated.
- Sharing a detector result with the class or with parents before you check the drafts.
- Pasting student work into a public AI tool to ask if it is AI-written. Follow your school's rules on student records first.
- Reading only the final file and skipping the history.
- Lowering a grade before the student has answered your questions.

## Where Lazur fits, and where it stops

Lazur can help with the mechanics. It explains what is on your screen when you hold ⌥ ⌘, and it points at buttons and menu items with a circle and a label. It guides you one click at a time, waiting for each click. It does not click through Pages or Google Docs for you. It does not decide whether a student cheated. That call stays with you and your school.

## What to do with the evidence this week

Pick one essay this week and run the checks in order. Start with the detector result, and treat it as a question. Then read the history, compare an earlier sample and hold a short talk. Write down what you saw and what the student said, with dates. If the evidence still points to a problem, use your school's academic integrity process.

Give the student a chance to respond before any grade changes. Your district's policy decides what happens after that. If you are unsure who to ask, start with your department head or an administrator before you act.

## With Lazur

Lazur is a voice assistant for the Mac that lives in the notch. Hold ⌥ ⌘ and ask; it sees your screen, answers out loud, points at buttons and walks you through apps. Hold fn to dictate. Free for 3 days, then $18/month: https://www.lazur.app/download

- “Where is the version history in this doc?”: Lazur circles Last edit at the top right and labels it, so you click it yourself. (Points at it)
- “Explain what I am looking at in this version history.”: Hold ⌥ ⌘ and ask. Lazur answers out loud about what is on screen. It does not decide who wrote the essay.
- “Where is Browse All Versions?”: Lazur draws a circle and a label around File > Revert To > Browse All Versions in Pages, so you find it without hunting.
- “Walk me through giving edit access in Google Docs.”: Lazur guides you one click at a time and waits for each click. You do the clicking and can stop at any step.
- “Draft a neutral note asking for the draft history.”: Lazur drafts text for you to read and change. You send the message yourself from Gmail.
- “Dictate my notes from the talk with the student.”: Hold fn to dictate clean text at the cursor in any app. Dictation is unlimited on both Pro and Power plans.
- “Find the official AI policy page for my district.”: Research runs in Lazur's own Chrome tabs and comes back with a summary and a report. Open the source page yourself before you quote it.

## A plan you can start this week

- **Day 1: Write the AI rule for this essay.** Name what counts as help, what does not, and what students must disclose. Write a few examples on each side. Post the rule with the assignment, where students see it before they start.
- **Day 2: Set checkpoints and access.** Ask for an outline, a draft and a final file as separate submissions. Share each Google Doc with edit access so the history is visible. Confirm that you can open each file's history before the first checkpoint is due.
- **Day 3: Take a baseline.** Save a short in-class paragraph from each student with the date on it. Note the vocabulary range and the usual errors. You will compare against this later, so keep it in your grading files.
- **Day 4: Check one flagged essay.** Run the signals in order: the score as a question, then the history, then the baseline. Write down dates and times as you go. Do not share the score yet.
- **Day 5: Hold a short talk.** Ask a few questions about the thesis, one source and one sentence the student would change. Keep the talk short and calm. Write down what the student says in their own words.
- **Day 6: Record the evidence and decide.** Record the evidence, the talk and the school process you will follow. If the evidence is still unclear, ask your department head before any grade changes.
- **Day 7: Update the rule.** Change one line of the AI rule based on what you learned this week. Share the change with students before the next assignment.

## FAQ

### Can an AI detector prove a student used AI?

No. A detector score is a signal, not proof. In a 2023 study, detectors flagged 61.22% of human-written TOEFL essays by non-native English writers as AI-generated. Use the score to decide where to look next, then check drafts and talk with the student.

### Where is version history in Pages on Mac?

Open the document and choose File > Revert To > Browse All Versions. Click Restore to make a version current, or Option-click Restore a Copy to open a copy in a new window. Work on the copy so the student's file stays as it is.

### Why can't I see version history in a Google Doc?

Google requires edit permission to browse earlier versions. Viewers and commenters cannot see the history. If you have edit access, click Last edit at the top right. If you only have view access, ask the student to share the document with edit access for this assignment.

### Can Lazur check the essay for me?

No. Lazur explains what is on screen when you hold ⌥ ⌘, points at buttons and menu items with a circle and a label, and guides you one click at a time. It does not click through Pages or Google Docs, and it does not decide whether a student cheated.

### Can I paste a student essay into a chatbot to ask if it is AI-written?

Check your school's rules on student records first. Your school may restrict what you can paste into outside tools. Even where pasting is allowed, a chatbot's answer is another guess, not evidence. Do not use it as the basis for an accusation.

### What if the student says they used a grammar checker?

Ask which tool they used and what it changed. If the student used Grammarly and your school allows it, the Authorship report may show typed, pasted and AI-generated text. A report exists only after someone generates a link and clicks Share.

### What if the evidence is still unclear?

Keep the case open and follow your school's process. Give the student a clear window to respond, in writing or in person. Record what you asked and when. Do not change a grade based on a score alone.

### Do students with accommodations face different checks?

Check the accommodations on file, and talk to your school's support staff, before you judge the work. Accommodations can change how text is produced. Your school's rules decide what counts as acceptable help.

## Sources

1. [GPT detectors are biased against non-native English writers (arXiv 2304.02819, HTML v3)](https://arxiv.org/html/2304.02819v3), arXiv, checked 2026-11-09
2. [GPT detectors are biased against non-native English writers (arXiv abstract)](https://arxiv.org/abs/2304.02819), arXiv, checked 2026-11-09
3. [AI Detectors Don't Work. Here's What to Do Instead](https://mitsloanedtech.mit.edu/ai/teach/ai-detectors-dont-work/), MIT Sloan Teaching & Learning Technologies, checked 2026-11-09
4. [Student, teacher AI use continued to climb in 2023-24 school year](https://www.k12dive.com/news/student-teacher-ai-use-schools-cdt/737335/), K-12 Dive, checked 2026-11-09
5. [Most Teachers Receive No Formal Guidance on AI Use](https://news.gallup.com/poll/710534/teachers-receive-no-formal-guidance.aspx), Gallup, checked 2026-11-09
6. [The AI Dividend: New Survey Shows AI Is Helping Teachers Reclaim Valuable Time](https://www.waltonfamilyfoundation.org/the-ai-dividend-new-survey-shows-ai-is-helping-teachers-reclaim-valuable-time), Walton Family Foundation, checked 2026-11-09
7. [Restore an earlier version of a document in Pages on Mac](https://support.apple.com/guide/pages/restore-an-earlier-version-of-a-document-tan7f1de6ec5/mac), Apple Support, checked 2026-11-09
8. [See version history in Google Docs](https://support.google.com/docs/answer/190843), Google Docs Editors Help, checked 2026-11-09
9. [Introducing Authorship](https://support.grammarly.com/hc/en-us/articles/29548735595405-Introducing-Authorship), Grammarly Support, checked 2026-11-09
10. [High School Teachers: Occupational Outlook Handbook](https://www.bls.gov/ooh/education-training-and-library/high-school-teachers.htm), U.S. Bureau of Labor Statistics, checked 2026-11-09