AI Detector Flagged My Thesis: What to Prepare

If an AI detector flagged my thesis, the meeting itself is not the hard part — the preparation is. Before you sit down with your committee, gather three things: your draft history, your supervisor correspondence, and a one-page written explanation of what the detector actually measured. That page does more work than anything you say out loud.

Key takeaways

• Bring evidence of process, not just a denial: dated drafts, version history, supervisor emails, notes.
• Write one page explaining what the detector measured — a probability estimate, not proof of authorship.
• Present your writing process as a timeline, so the committee can see how the thesis was built.
• Never argue that detectors are useless. Argue that this score, on this text, is a false positive.
• If the committee asks for a formal response, follow their format and deadline exactly.

What should I gather before the committee meeting?

The committee is not really asking “did you use AI?” They are asking “can you show us how this text came to exist?” Your job is to answer that second question with documents.

Draft history. Pull every version you still have. Chapter files, dated backups, cloud document revision history, even old email attachments you sent yourself. A file named chapter3_v4_2025-03-11.docx is worth more than a paragraph of explanation. If you write in a cloud editor, export the revision history as a PDF — it shows timestamps and edit patterns that a final file cannot.

Version control. If you use Git, Overleaf, or a university document system, export the commit log or history view. If you use plain folders, build a timeline: which chapter existed when, what changed, and why. Handwritten notes, index cards, and annotated printouts are legitimate evidence. Bring them.

Supervisor correspondence. Every email where you sent a draft and got comments back is a dated record that the text existed and was being revised over time. Search your inbox for the chapter names. Screenshot or export the threads, including the attachments.

Reading and source notes. Annotated PDFs, library loan records, Zotero or Mendeley libraries, interview transcripts, lab notebooks. These show the material the thesis was built from.

Put all of it in one folder, in date order, and bring a printed index page. You want the committee to see a project that grew over months, not a document that appeared fully formed.

How do I explain what the detector measured in one page?

This is the part most students get wrong. They walk in and say “the detector is wrong.” That is a conclusion. What the committee needs is a description of the instrument.

Write one page with four short sections:

1. What the tool claims to do. It estimates the probability that a passage was generated by a language model, based on statistical patterns in the text.
2. What the score is. It is a probability estimate, not a finding of fact. A high score means the text resembles patterns the model associates with machine output. It does not identify an author, and it cannot see your process.
3. Why false positives happen. Formal academic prose, dense citation, repetitive terminology, non-native English phrasing, and heavy editing can all push a score up. Our own September 2026 measurement illustrates how noisy this is: across 46 hand-written DeAIze guides, the mean AI score was 27% and the median 25%, with a range from 0% to 55% — all of it written by people.
4. What the score cannot tell you. It cannot distinguish “written by AI” from “written in a style that resembles AI.” Those are different claims, and only the first one is misconduct.

Keep it to one side of A4. Number the sections. Hand a copy to each person at the table. If you would rather understand the underlying signals before you write it, our guide to how AI detection works covers the mechanics without the panic.

If you want to see which specific sentences in your own draft carry the signal, DeAIze’s AI detector scores the whole document and highlights each sentence by AI probability, so you can walk into the room knowing exactly which lines are in question.

How do I present my writing process without sounding defensive?

Open with the process, not the accusation. A short spoken version of your folder index works well:

“I started chapter one in October, sent it to my supervisor in November, revised it in January after her comments, and the version you have is the fourth draft. Here is the timeline, and here are the emails.”

That is a factual statement about a project. Nobody has to be accused of anything for it to be true.

The trap is tone. If you spend ten minutes attacking the detector, you look like someone with something to hide. If you spend ten minutes calmly walking through dates and documents, you look like a researcher. The evidence does the arguing; you just present it.

Three practical rules:

• Answer the question asked. If they ask when you wrote a section, give the date. Do not pivot to a speech.
• Concede the small stuff. If you used a grammar tool, an AI-assisted literature search, or ChatGPT to brainstorm, say so — check your university’s policy first, and be honest about what is permitted. Concealing a permitted tool is worse than the tool itself.
• Keep the ideas yours. The argument, data, and conclusions have to be your own work. Detection is a sideshow; authorship is the actual subject of the meeting.

What not to say when your thesis is flagged

Some sentences make the conversation harder. Avoid these:

Don’t say	Why it backfires	Say instead
“Detectors are all fake.”	Attacks the tool instead of addressing the text	“This score is a probability estimate, and here is why it misreads this passage.”
“Everyone uses AI.”	Sounds like a confession with company	“Here is what I used and what my policy allows.”
“I don’t know how that sentence got there.”	Suggests you did not write it	“That paragraph came from my January revision — here is the earlier version.”
“My writing has always been like this.”	Unverifiable	“Here are three earlier papers with the same style.”
“I’ll just rewrite it.”	Implies the original was a problem	“I can show the revision history for that section.”

Notice the pattern: every good answer points at a document. Every bad answer points at a feeling.

If you have already been through a false positive in another context, the false positive playbook walks through the same evidence-gathering steps in more detail.

How do I prepare for a formal response if the committee requires one?

Some committees escalate. You may be asked for a written statement, a meeting with an academic integrity officer, or a response to a formal allegation. The process varies by institution, so read your university’s policy before you write anything.

What usually helps:

1. Ask for the specifics in writing. Which passages, which tool, which score, which policy clause. You cannot respond to a vague allegation.
2. Get the deadline and the format. Word limit, submission channel, who reads it.
3. Reuse your one-pager. Your explanation of what the detector measured belongs in the formal response almost unchanged.
4. Attach the evidence index. Draft timeline, correspondence, notes. Reference each item by number.
5. Ask whether you may bring a supporter. Many institutions allow a supervisor, advisor, or student union representative to attend.
6. Keep a copy of everything you submit. Date it.

If your department is still deciding how to treat AI-assisted work in general, what your university actually allows is worth reading before you draft your response — policies differ sharply on permitted tools, and quoting your own policy accurately is far stronger than arguing from principle.

What if the flagged passages are genuinely machine-assisted?

Be honest with yourself here, because the committee will get there anyway.

If you used a model to draft or rewrite passages and your policy does not permit it, the meeting is not about the detector. It is about the draft. In that case, the useful preparation is different: identify exactly which passages were assisted, understand what the policy says about that, and decide with your supervisor how to address it. Rewriting the flagged sections so they are genuinely yours — your argument, your phrasing, your evidence — is a legitimate move if your policy allows the revision and you disclose it. Our guide to cleaning up machine-made patterns is written for that situation.

What you should not do is run the thesis through a tool and submit the output as if nothing happened. That converts a policy question into a honesty question, which is much worse.

Does the score itself prove anything?

No. This is worth saying plainly, because the fear in these meetings is usually disproportionate to what the evidence supports.

A detector score is a probability estimate about text patterns. It is not a statement about authorship, intent, or process. Two tools given the same paragraph routinely disagree, which is why two detectors can return different verdicts on identical writing. Our own measurement found unedited model output scoring a mean of 35% at a 30% flag threshold — below the human sample’s upper range in places, and overlapping it almost entirely. A tool that cannot cleanly separate its own generated text from human guides is not a tool that can identify who wrote your chapter three.

That does not mean you walk in and dismiss the report. It means you treat the score as one noisy input among many, and you bring the inputs it cannot see: your drafts, your emails, your notes, your process.

A five-day preparation plan

If the meeting is soon, work backwards.

Day	Task
1	Export revision history and build the draft timeline.
2	Search email for supervisor threads and export them with attachments.
3	Assemble reading notes, annotated PDFs, lab or interview records.
4	Write the one-page explanation of what the detector measured.
5	Rehearse the two-minute process summary out loud. Print the evidence index.

If you have more time, use it to read your university’s policy properly and to re-read the flagged passages so you can talk about them specifically. Nothing lands better than a student who can explain why a particular sentence is phrased the way it is.

Ready to see which sentences read machine-made?

The most useful thing you can do before the meeting is know your own document. DeAIze’s AI detector scores the whole thesis and highlights each sentence by AI probability, so you can see exactly which lines a committee member is likely to ask about — and prepare a specific answer for each one instead of a general defence. If the flagged sentences are genuinely your own writing, the draft editor lets you review and revise them line by line without flattening your voice, and pricing is pay-as-you-go with credits that never expire. You can also create a free account and start with 200 words, no credit card. Keep the ideas, data, and conclusions your own — the tools are for cleaning up your own prose, not for rewriting someone else’s work.

What if the flagged passages are in a chapter you wrote months ago?

Time is your friend here, and most students forget to use it. A passage written in October and revised in January has a paper trail that a passage written last week does not. Pull the earliest version you can find and compare it against the final text. If the sentence that triggered the flag existed in near-identical form six months before the meeting, say so and show the file date.

A short checklist for old chapters:

1. Find the earliest surviving draft of that chapter — email attachment, cloud revision, printed copy.
2. Note the date and where it came from.
3. Highlight the flagged sentence in both versions.
4. Write one line explaining what changed between them and why.

If the flagged sentence is unchanged across drafts, that is strong evidence of consistent authorship. If it changed, the revision history shows who changed it and when.

One caution: do not assume an old file date settles the question. It shows the text existed, not who typed it. Pair it with correspondence — a supervisor’s comments on that chapter, a peer review note, a submission receipt. For how a detector reads a single sentence in the first place, our DeAIze review walks through what the report actually shows, sentence by sentence, so you can match your evidence to the specific lines under question.

If you have no old drafts at all — common for students who work in one long file — say that plainly and lean on correspondence and notes instead. An honest gap is better than a reconstructed timeline.