How to Prove You Wrote It Yourself: An Evidence Guide for AI Accusations

Start Here: Evidence Beats Argument

Most people accused of AI use respond by disputing the detector. That is the weaker move. Detector accuracy is contested, technical, and easy for a reviewer to wave away.

Evidence that you did the work is not contested. It is either there or it is not.

Approach	Persuasiveness	Why
"The detector is unreliable"	Low–moderate	True, but reads as deflection
"I didn't use AI"	Low	Unverifiable assertion
"Here is the document being written over four days"	High	Shows the work directly
"Here are my notes, sources and outline"	High	Hard to produce retrospectively
"Ask me anything about this argument"	High	Immediate and hard to fake

Lead with the bottom three. Use the top one only as supporting context.

---

The Evidence Hierarchy

Tier 1 — Timestamped process records. These are the strongest, because they show the document existing in incomplete states over time.

• Google Docs version history — File → Version history → See version history. Shows every editing session with timestamps.
• Microsoft Word version history — available for files saved to OneDrive or SharePoint.
• Track Changes, if you had it on.
• Cloud backup snapshots — Dropbox, iCloud and OneDrive all keep prior file versions.

Tier 2 — Working materials. These show thinking that precedes writing.

• Research notes, in any form
• Outlines and structure plans
• Annotated PDFs or highlighted readings
• Reference lists you built as you went
• Photographs of handwritten notes, with file dates

Tier 3 — Circumstantial trail. Weaker alone, useful in combination.

• Browser history showing research sessions
• Library borrowing or database access logs
• Emails or messages discussing the assignment
• Supervisor or tutor conversations about your progress

Tier 4 — Demonstrated understanding. Often decisive in person.

• Explaining why you structured the argument as you did
• Discussing sources you cited and why you chose them
• Describing what you cut and why

---

What To Do in the First 48 Hours

1. Do not delete anything. Not drafts, not notes, not browser history. Even material that feels irrelevant may establish a timeline.

2. Export your version history immediately. Cloud version history is not kept forever. Take screenshots or export while it exists.

3. Ask for the specific allegation in writing. You are entitled to know which tool produced what score, against which policy. "It was flagged" is not an allegation you can answer.

4. Ask for the policy text. Many institutions have vague AI policies. If the rule you allegedly broke is not clearly stated, that matters.

5. Do not admit to something you did not do to make it go away. Pressure to resolve quickly is common. An admission is far harder to undo than an appeal is to file.

---

Building Your Response

A response that works has four parts, in this order:

Part	Content	Length
1. Direct statement	"I wrote this work myself. Here is the evidence."	1–2 sentences
2. Process evidence	Version history, drafts, notes, with dates	The bulk of it
3. Context on the tool	Published false positive rates, vendor guidance	1 short paragraph
4. Offer to discuss	"I am happy to talk through the argument"	1 sentence

For part 3, two facts are worth knowing:

• Turnitin's own guidance states the AI writing score should not be used as the sole basis for action against a student.
• Liang et al. (2023), published in Patterns, found a 61.22% false positive rate when seven commercial detectors were run on TOEFL essays by non-native English speakers — all human-written. If English is not your first language, this is directly relevant. See our full breakdown of AI detectors and non-native English writers.

Keep part 3 short. It is context, not your case.

---

If English Is Not Your First Language

You are at meaningfully higher risk, and the research says so. The same study found 97.80% of TOEFL essays were flagged by at least one detector out of seven.

Say this explicitly in your response, with the citation. It reframes the score from evidence about you into a known limitation of the tool affecting a specific group — which is exactly what it is.

---

Protecting Yourself Going Forward

None of this requires you to change how you write. It requires a record.

Habit	Time cost	Value if accused
Draft in Google Docs or Word online	Zero	Automatic version history — the single best protection
Keep an outline file	5 minutes	Shows structure preceded prose
Save research notes in one place	Ongoing, small	Shows the reading behind the writing
Write across multiple sessions	None — most people do anyway	Timeline is far more credible than one long session
Keep the messy draft	Zero	Imperfection is strong evidence of human process

That last row deserves emphasis. People instinctively delete bad drafts. Do not. A document that starts disorganised and improves is close to unfakeable, and it is exactly what a real writing process looks like.

---

A Note on Tools

If you use an editing tool — a grammar checker, a translator, an AI humanizer — that does not make the work not yours, but it may change what you have to disclose. Policies differ enormously and many are unclear.

Ask what your institution actually requires, in writing, before you submit. That question protects you far more reliably than any tool does.

Related: why writing gets flagged as AI · what to do about a detector false positive