You wrote the paper yourself. Then a message arrives saying an AI-detection report flagged part of it. (Illustrative example: "Your submission returned a high AI-writing score. Can we meet to discuss it?")

If that is where you are, take a breath. A detector score is a statistical estimate, not a finding about who wrote your paper, and even the vendors say it can be wrong. This guide covers what these tools look at, why careful human writing can be misread, and how to respond with evidence instead of panic.

One note up front: this is general information, not legal or academic-integrity advice. Your institution's policy and process govern what happens next.

What AI detectors look at

Commercial detectors do not publish full details of their models, and they differ from one another. Much of the early public explanation of AI detection centered on two ideas:

Perplexity: how predictable your word choices are to a language model. Text a model finds very predictable looks more "machine-like" by this measure. Clear, conventional, well-edited prose can also be predictable.
Burstiness: how much sentence length and structure vary across a passage. Very even rhythm looks more machine-like by this measure, even though plenty of careful human writing is even.

Many tools now go beyond those two signals. GPTZero, for example, says its current model relies on deep learning and treats perplexity and burstiness as only some of its indicators. But every approach shares one limitation: the tool infers authorship from patterns in the finished text. It cannot see how the text was written. For a deeper explanation, see what perplexity and burstiness mean.

Why careful human writing can get flagged

The best-known evidence is a 2023 study by Stanford researchers Liang and colleagues, published in the journal Patterns. They ran 91 human-written TOEFL essays by non-native English writers through seven AI detectors available at the time. On average, the detectors labeled 61% of those essays as AI-generated. Essays by US eighth-graders were classified far more accurately, with about 5% misclassified on average (Liang et al., 2023).

The authors tie this to perplexity: writers who rely on common, safe vocabulary produce more predictable text. When the researchers had a chatbot enrich the vocabulary of the same essays, the average false-positive rate fell sharply. The detectors were reacting to word choice, not to who wrote the essay.

That study tested specific 2023 tools on a small set of essays, so its numbers are not the error rate of any detector you face today. But the mechanism it describes is a reasonable caution for other writing that tends toward predictable wording or even rhythm, for example:

Writers working in a second language, who often favor vocabulary they are confident in.
Formal or technical writers, whose fields use standard, repeated terminology.
Heavily edited prose, where revision has smoothed out the unexpected turns.

Detector makers acknowledge errors too. OpenAI withdrew its own AI text classifier in July 2023, citing its low rate of accuracy (OpenAI). Turnitin's guidance says its AI writing detection may misidentify both human and AI-generated text and should not be used as the sole basis for adverse action against a student (Turnitin guides).

What to do if you are flagged

Step 1: Read the message carefully. Note what was flagged, which tool was used, and what your instructor is asking for. A request to talk is not a finding of misconduct.

Step 2: Gather your version history. Revision history shows a document developing over time, which a single score cannot show.

Google Docs: File → Version history → See version history. Screenshot early and late versions with their timestamps.
Microsoft Word: Version history is available for files saved to OneDrive or SharePoint (Microsoft support). If you worked offline, gather dated backup copies instead.

Step 3: Gather other process evidence. Notes, outlines, annotated sources, earlier drafts, research searches, and messages about the assignment can all help show how the work came together. They are supporting evidence, not a guarantee of any particular outcome.

Step 4: Ask for a meeting in writing. A short, polite email works: "I'd like to discuss the report. I wrote this paper myself and can share my version history and notes. Could we meet this week?" A written request keeps a clear record.

Step 5: Read your institution's policy. Find the academic-integrity policy and the process for responding to a concern, including any right to an advisor, to respond in writing, or to appeal. If you are unsure, ask your academic advisor or the dean of students office what support is available. If a score is being treated as proof on its own, the vendor guidance above is worth pointing to.

Step 6: Be ready to talk about your work. Explaining why you made specific choices in your argument, sources, and structure is strong evidence that the thinking is yours.

Habits for every paper going forward

Write where history is kept. Draft in a tool that saves version history automatically, such as Google Docs, or Word with the file saved to OneDrive, so the record exists if you ever need it.

Keep your process trail. Save outlines, notes, and annotated sources in the same folder as your drafts. It takes seconds, and it is the same kind of record researchers and journalists keep.

Follow your course's AI policy. If AI tools are allowed for brainstorming, outlining, or editing, disclose them the way your instructor or institution requires. If the policy is unclear, ask before you use one.

Editing tools, used within the rules

Editing tools can help you revise your own writing for clarity. None of them can prove authorship, and none can promise any detector result.

Spelling and grammar checks built into Google Docs and Word, or a dedicated checker such as Grammarly, can catch errors. Review every suggestion.
Hemingway Editor highlights long or complex sentences and passive voice, and offers a free web version.
HumaText can propose a phrasing pass on a short section of a draft, where your course or school policy permits AI-assisted editing. Review every change against your sources and intended meaning. It is free to try without an account.

The longer view

Detection tools estimate whether text resembles machine-generated text. That is a different question from who wrote it, and the published research shows the two can diverge, especially for writers whose prose is predictable for innocent reasons.

The right response is not to write worse. It is to keep writing carefully, keep a record of your process, and bring that record to any conversation. A score can start a conversation. Your drafts, your notes, and your ability to explain your work are what answer it.

Why AI Detectors Can Flag Your Own Writing, and What to Do About It