My professor thinks I used AI
The conversation goes better when you know what the number is, what it is not, and which properties of your own writing produced it.
Somebody has looked at a percentage and formed a view. The percentage came from a classifier with a published error rate, and that error rate is the most useful thing you can bring to the discussion.
What follows procedurally is set by your institution and varies a great deal, so this page stays on the part that is the same everywhere: what the number actually measures.
Three facts, all from the tool vendors
Wrong on human writing 4% of the time. Turnitin publishes this. It is not a rare event, it is a designed-in rate, and at any real volume it happens constantly.
Below 20% is thinner. Turnitin states that documents scoring under 20% show a higher incidence of wrong flags.
Wrong flags cluster. Turnitin's analysis found 54% of falsely flagged sentences sit directly beside a genuinely flagged one and 26% two sentences away. Four in five land next to a real detection rather than scattering.
What the indicator is and is not
It is a probability that your writing has properties characteristic of generated text.
It is not a match against a source. Nothing was found, so there is nothing to click through to. This is the difference between the AI indicator and the similarity score sitting next to it on the same report, and the two get conflated constantly because they are presented identically. See how Turnitin AI detection works.
Turnitin's own documentation states the indicator should not be the sole basis for action.
Move the conversation to the specific
A general dispute about whether detectors work goes nowhere. A specific one about your document goes somewhere, because the properties are countable.
Run the document through a second detector that scores per paragraph. Then you can say which sections read as flat and why:
A methods or procedure section is formulaic by design. Narrow vocabulary, conventional constructions, deliberately repetitive. It reads machine-like to a classifier because it is regular, and it is regular because that is how the genre works.
A heavily revised section has had its variation edited out. Our corpus puts human writing at a sentence-length standard deviation of 8.7 words against 6.4 for generated text, and every revision pass moves a document toward the lower number.
Writing in a second language shows elevated wrong-flag rates in published evaluations, repeatedly. See why non-native English writing gets flagged.
Being able to point at a paragraph and name the property is a different conversation from asserting you did not do it.
Bring the record
Version history is the strongest thing you have, and it is usually already sitting there in Google Docs or OneDrive without you having done anything. Drafts, supervisor comments, notes, reading history. Full list in how to show you wrote your essay.
Before the next one
The reason this situation is uncomfortable is that everything useful happens before submission and you did not have the information then. That part is fixable and costs a few minutes.
Ours is free and unlimited with no account, takes .pdf and .docx, scores every paragraph separately, publishes its wrong-flag rate at 0.4%, and never stores your text.