How to prove you didn’t use AI on an essay

What actually counts as evidence

An accusation of AI use usually arrives as a number: a detector score, sometimes a screenshot of a highlighted paragraph. The way out is not a better number. It is a record of the work. Integrity offices and instructors are used to weighing drafts, notes, sources, and a conversation about the paper. Give them those, in that order, and let the score sit next to them.

Here is what tends to be accepted, roughly in order of weight.

Drafts with timestamps

Every file you saved on the way to the final: the outline, the first draft, the version you sent to a friend, the one you cut in half. Your file system keeps creation and modification times. Email attachments carry dates. A draft you emailed to yourself or a classmate before the due date is dated by someone else’s mail server, which is hard to argue with. Gather all of it before you reply to anyone.

Version history, and its limits

If you wrote in Google Docs or Word, the version history shows the document growing over time. Instructors accept it often, and it belongs in your packet. Know its limits before you rely on it alone. Google Docs and Word both record document states, not keystrokes, and neither can tell typed text from pasted text. Word’s Track Changes can be switched off by the author. Tools sold as “human auto typers” replay pasted text into Docs or Word as keystrokes to fabricate a natural-looking history, and some integrity offices know this and discount history accordingly. Bring it anyway. The page on version history shows how to export it and how to present it.

Source: Microsoft Learn Q&A on Word version history; phrasly.ai human auto typer page; checked 2026-09-08.

Notes, sources, and reading

Your annotated readings, highlighted PDFs, library search history, bookmarks, the notes app entry from the night you got the idea, a message to a classmate about the thesis. Anything that shows the argument existed in your head before it existed in the document. This material is easy to overlook and surprisingly persuasive, because a chatbot does not leave it behind.

The ability to explain it out loud

This is the evidence instructors trust most, and the one no tool can give you. Be ready to walk through the paper without looking at it: why you picked the topic, what the main claim is, which paragraph is weakest and why, which source you almost used and cut. If you wrote it, this is easy. Ask for the chance to do it, in person or on a call.

A writing-process record

Some courses have students write in an editor that records the session itself: every keystroke, revision, pause and paste, with server-side timestamps. That kind of record shows what a document history cannot. It shows whether a sentence was typed or pasted, how long you sat before starting the second paragraph, which sentences were rewritten, and the order in which the argument formed. If your course used one, ask for your session to be reviewed. If it did not, you cannot build one after the fact, and you should not try; a fabricated history reads as unrecorded, not as writing.

Why a second detector run proves nothing

The first instinct is to paste the essay into another detector and screenshot the low score. Do not lean on it. A detector reads the finished prose and guesses. Two guesses that disagree do not tell anyone who wrote the paper. What a disagreement can show is narrower: that the tools do not agree with each other, which is a reason to treat the first score as a probability and not a finding.

The published research says the same thing, and you can cite it. Stanford researchers (Liang and colleagues, Patterns, 2023) found that GPT detectors flagged 61.3 percent of TOEFL essays by non-native English writers as AI-generated. Common Sense Media (2024) reported that 20 percent of Black teens said they had been falsely accused of AI use, against 7 percent of white teens. Turnitin publishes a page titled “Understanding false positives within our AI writing detection capabilities” and states a document-level false positive rate under 1 percent for documents with more than 20 percent AI writing. One in a hundred, across a whole campus, is a lot of students. The false positive rates page collects every published number.

Source: Liang, Yuksekgonul, Mao, Wu, Zou, Patterns, 2023; Common Sense Media, 2024; turnitin.com false positives page; checked 2026-09-08.

What happened in Newby v. Adelphi

In early 2026 the New York Supreme Court, Nassau County, decided Newby v. Adelphi University (2026 NY Slip Op 26021). The student’s paper had been flagged by Turnitin at 100 percent AI. Grammarly and ZeroGPT read the same paper as human. The court found the university’s finding “without valid basis” and ordered the record expunged. Two things are worth taking from it. The court did not rule that the other detectors were right; it ruled that a finding resting on the score could not stand. And the student had to go to court to get there. Most students will not, which is why the evidence and the conversation matter so much earlier in the process. The lawsuit tracker lists the other cases so far.

Source: Justia, 2026 NY Slip Op 26021; checked 2026-09-08.

How to write the email

Keep it short, factual and polite. Send it to the instructor first unless the matter has already moved to an integrity office. Do not argue about the detector in the email. Ask for the process and offer the evidence.

Dear Professor [name], I received your message about my [assignment] and the AI detection result. I wrote this paper myself and I would like to show you how it came together. I have my outline, two earlier drafts with file timestamps, my notes on the sources, and the version history from Google Docs. I would also welcome the chance to talk through the argument with you in person or on a call. Could you tell me which policy applies and what the next step is? I can send the materials today.

Sample wording. Change it to match your course and your evidence.

What to do next

1. Save everything now: drafts, notes, the document with its history, screenshots of messages. Copy it to a second place.
2. Ask, in writing, for the specific evidence against you and the policy that applies. A score is not a policy.
3. Send the email above with your materials attached or linked, and ask for a meeting.
4. Prepare to talk through the paper: thesis, structure, sources, what you would change.
5. If the finding stands, ask for the appeal route and contact your student advocacy office or ombuds. The step-by-step page covers the hearing.

What not to do

• Do not admit to something you did not do because it seems like the fastest way out. Findings follow you.
• Do not delete or edit anything. An edited history looks worse than an incomplete one.
• Do not build your case on a second detector score. Use the research to question the first score, then move to your evidence.
• Do not try to manufacture a history with an auto typer. It is detectable and it turns a false accusation into a true one.

The limits of any record

A process record, a version history, a folder of drafts: each shows how text entered the document. None shows whether the ideas were yours. That is why the conversation matters more than the packet. The evidence earns you a fair conversation. The conversation is where you show that you wrote it. Ask for that meeting, bring what you have, and explain the paper the way you would explain it to a friend.