Copilot AI Detector

Looking for a Copilot AI detector or Copilot AI checker? Paste 100 words or more of the text you want to review into the free checker above, whether it is an email drafted in Outlook, a report from Word or an answer from the Copilot app. V4, our current engine, scores it for AI-writing patterns from 1 to 99 and returns a stability range and a verdict: Likely human-written, Uncertain or Likely AI-written. It is free, there is no sign-up and the text is not kept.

One thing makes Copilot different from the other guides on this site. Copilot is a Microsoft product, not one model family, so there is no Copilot-only test figure. What we can tell you is which of the models behind Copilot V4 was trained on, how well it caught their writing in our sealed test, and how to read a score on a Copilot draft.

Is there a Copilot AI detector?

Yes, in the sense that matters for most people: the checker above measures AI-writing patterns in any English text, including text drafted with Copilot. No tool can confirm that Copilot was used, though. V4 never names a model or a product, and a Copilot answer can come from more than one underlying model.

• It can score English text of 100 words or more and place it in one of three verdicts.
• It was trained on writing from two of the model families Copilot can use: OpenAI’s GPT models, as in ChatGPT, and Anthropic’s Claude.
• It cannot say whether Copilot, ChatGPT, Claude or another tool produced a passage.
• It has not been tested on Copilot’s own output or on Microsoft’s own models.

What the Copilot AI checker measures

V4 is a fine-tuned transformer text classifier. It reads a passage in overlapping windows, averages them into one internal value and maps that value to a score from 1 to 99. It was trained once and frozen, on texts from nine model families and on a large body of human writing dated 2021 or earlier, so it recognises how current AI writing differs from human writing in general rather than one product’s habits.

It does not look at document metadata, tracked changes or who was signed in when the text was written, and it does not compare the text with other files. It scores the words you paste, as one passage. How the detector works covers the engine in more detail.

The score is not the share of the text written by AI, and it is not a probability. A result alone does not prove who wrote a text.

Which models Copilot uses, and what V4 was trained on

Depending on the app, the plan and the settings, Copilot can answer with OpenAI models, Anthropic’s Claude models or Microsoft’s own models, inside Microsoft’s own prompting and product layer. V4’s training data included writing from the first two of these families, and our sealed test of 17 October 2026 measured both on texts the engine had never seen:

• OpenAI’s models, tested through ChatGPT. 525 of 525 ChatGPT texts of 150 words or more were caught (100.00%, Wilson 95% range 99.27 to 100.00). The ChatGPT detector guide has the versions and the figures by length.
• Anthropic’s Claude. 1,045 of 1,048 Claude texts of 150 words or more were caught (99.71%). The Claude AI detector guide has the details.
• Microsoft’s own models. These were not in training and were not measured, so no figure exists for them.

Those results describe what each model writes in its own app or API. They are the closest evidence we have for Copilot, but they are not a Copilot result, because Copilot adds its own instructions, document context and formatting on top.

Checking Copilot drafts from Word, Outlook and Teams

Much Copilot text never leaves a document. Someone asks Copilot in Word to draft a section, accepts it, then rewrites a sentence here and there; an Outlook reply gets a quick polish before it is sent. By the time you see it, the text is a mix of model wording and human edits. That mixed text is exactly what our tests have not measured for any family, and edits usually pull the score down.

Two practical tips help. If you have access to the document’s version history, check the earliest version that contains the passage, not only the final one. And check sections separately as well as the whole document: a single Copilot-drafted section can be diluted by pages of human writing around it.

How to check a Copilot draft

1. Pick the passage you have a question about, ideally as it first appeared.
2. Paste at least 100 words into the checker above, or 150 words or more for a steadier reading and fewer false flags.
3. Run the check and read the verdict, the score and the stability range together.
4. Read the evidence line, which shows how often human writing of the same length reached that zone in our sealed test.
5. Compare and record. Put up to 3 passages side by side in compare mode, and save the result as a PDF if you need to keep it.

Reading the result on Copilot text

65 is the AI line at every length. Scores from 1 to 34 read as Likely human-written, 35 to 64 as Uncertain and 65 to 99 as Likely AI-written, with a Moderate signal chip from 65 to 84 and a Strong signal chip from 85 to 99. Each length band has its own frozen internal threshold, mapped so that 65 always sits on it, and the stability range shows how far the score moves when the source families behind the score map are resampled.

On Copilot text, read a high score as a sign that the passage still carries a lot of unedited AI wording. Read a low or Uncertain score as inconclusive, because edited Copilot drafts are the mixed case nobody has measured. Every line of the result card is explained in understanding results.

What we did and did not test

The sealed test measured ChatGPT texts produced in the ChatGPT app and Claude texts produced through Anthropic’s API and Claude Code, alongside the other model families, on 17 October 2026. Copilot’s own output, with its workplace prompting, document context and the edits people make in Word or Outlook, was not generated or scored. No text was produced through Copilot, in Microsoft 365 or anywhere else, so no recall figure for Copilot is published here. The evidence page lists everything that was measured, and the methodology explains how.

False flags and edited text

A false positive is human writing that reaches 65. In the sealed test that happened to 4 of 2,863 human texts of 150 words or more, about 1 in 700, and to 11 of 1,136 texts of 100 to 149 words, about 1 in 100. Learners of English were flagged about 3 times as often as native writers, and formal speeches reached the line in 17 of 931 cases. Formal workplace writing can look regular in the same way, so take care with policy documents and official letters.

If a document of yours was flagged, the version history in your editor is strong evidence of how it was written. What to do if your text is flagged explains how to use it.

Using a Copilot checker at work and at school

Many workplaces now allow or even encourage Copilot, so a high score may describe something entirely permitted. Check the policy first, then use the result to ask how a text was produced. Never discipline, grade or reject anyone on a score alone, and weigh drafts, history and a conversation with the writer above any number.

Copilot AI detector FAQ

Other AI model guides

For the families Copilot can use, see the ChatGPT detector and the Claude AI detector. For every other assistant we tested, including Gemini and DeepSeek, see all model guides.

Tested and written by the AI Detector Checker team. Last tested: 17 October 2026 (ChatGPT and Claude families; Copilot itself not tested).