I searched Turnitin's detector FAQ for "Copilot". Zero hits. The list it does publish is more interesting.

Copilot appears zero times in Turnitin's AI writing detection FAQ. I searched the rendered article text, all 31,014 characters of it, and got nothing. To make sure the search was actually working, I ran the same counter against words I could see on the page: Gemini 14, GPT 34, Claude 14. A deliberately impossible control string returned 0, as it should.

So the honest answer to "does Turnitin detect Copilot" is that the page which answers this question does not use the word. That is not the same as "no", and the reason why is the most useful thing on it.

The list names builds, not products

The published list is introduced with a verb worth copying exactly — can detect content from — and then runs 32 entries. Today, in order:

• 11 GPT builds (GPT-4o through GPT-5.4-pro), plus o1-mini
• 7 Gemini builds (Gemini-1.0-pro, Gemini-1.5-pro, Gemini-2.5-pro, Gemini 2.5 Flash, Gemini-3-flash-preview, Gemini-3-pro-preview, Gemini-3.1-pro-preview) — two of those render with a stray space in the middle on the live page, which is its own small signal about how the list is maintained
• 7 Claude builds
• 2 LLaMA, 1 Mistral, 1 Deepseek, 1 Nova, 1 Grok

Then it closes with a clause that does more work than the 32 entries in front of it:

"and tools based on these LLMs as well."

"We will continue to expand our detection capabilities to other models in the future."

That is a model-level list with a product-level catch-all bolted on the end. Which explains the zero hits: Copilot, like most things people actually type into, is a product surface sitting on top of models that may change under it without the label changing. The list does not enumerate surfaces. It names builds and then says "and whatever is built on these."

So the question "is my assistant covered" decomposes into a routing question the vendor's list cannot answer and you probably cannot either: which build served your text today. Gemini is a different case — seven of its builds are on there by name, so there is no inference step at all.

The list appears twice on the page, and the two copies disagree

This is the part I would not have caught by reading. The same list is printed in two places, once in the mechanism section and once under Scope of detection, and if you diff them entry by entry, 30 of the 32 match and two do not.

Entry	First copy	Second copy
o1-mini	released 2024-09	released 2024-12
Claude Sonnet-4.6	released 2026-02	released 2026-05

Both copies were on the page today. I do not know which date is right for either, and that is exactly my point: do not quote the release dates from this list. The model names are one assertion and the parenthetical dates are another, and the page itself does not agree with itself on the second one. Cite the names, skip the dates.

Three sentences that change what a score means

The rest of the page is worth reading precisely because it undercuts the confidence people attach to the number.

On the trade-off the classifier is tuned to:

"In order to maintain this low rate of 1% for false positives, there is a chance that we might miss some AI written text in a document."

That is a deliberate, documented bias toward not accusing people. The cost of it is on the other side of the ledger, and the vendor writes it down.

On the target moving:

"As we iterate and develop our model further to better detect newer LLMs, it is likely that our detection capabilities will also change, affecting the AI percentage."

And immediately after, the bit with real operational consequences:

"However, for a submitted document, the AI percentage will change only if it's re-submitted again to be processed."

Read those two together and you get something closer to a cache-invalidation rule than a grade. A score is a snapshot of the classifier as of the moment of submission. Nothing recomputes it. Two reports on the same document from different months are two different measurements, and comparing the two numbers without noticing that is a category error people make constantly.

The bit that is actually useful

• If you are asking "am I covered", you are asking the wrong side of it. The list names builds; you interact with products. Establishing which build served you is often not possible from where you sit.
• When you cite this list, cite the names and the catch-all clause, not the dates. The dates disagree with themselves on the same page.
• If you are comparing two reports, compare the reports, not the two percentages. Different submission dates can mean different classifier versions, and the earlier number was never updated.
• Screenshot anything you are going to rely on, with the URL visible. The page I read today carried an Updated 17 days ago stamp. It moves.

If you are holding a report and want to work on the passages it marked, HumanPen takes the exported Turnitin or iThenticate AI report, matches the flagged passages back to your document, and rewrites only the ones you confirm. Nothing outside that scope is touched, and it is designed to preserve citations, tables and formatting. Eligible passages can be re-run at no charge. I work on it: humanpen.net/humanize.

All quotes and counts above are from Turnitin's AI writing detection capabilities FAQ, read in a browser on 15 September 2026.