Model guide · reviewed 5 September 2026
Microsoft Copilot Detector
The tool below is the same detector as on the homepage.
English · 100+ wordsengine v3
Paste text to check
V3 uses the same trained core as V2 with new length-based thresholds and a 0-100 display map. How it works →
What the evidence shows
No Microsoft Copilot Detector-specific figure exists: the sealed test covered two current OpenAI models and one unseen model lineage. Human text from unseen sources received a wrong AI signal about 2 in every 1,000 times at 150 words or more. The detector never identifies which model wrote a text; this guide describes writing patterns to review, not a measured detection rate. See the evidence page.
Important: The tool does not identify the source model. Model version, prompt, settings, language, task, and human editing can make outputs from different systems overlap substantially.
Microsoft 365 Copilot writes inside the documents people already work in: a Word report, an Outlook reply, the summary at the top of a Teams recap. That is what makes reviewing it different from reviewing a chatbot answer pasted into a browser. The text arrives already formatted, already in the house style, and often already partly edited by the person whose name is on the file. This guide describes what that output tends to look like, what AI Detector Checker actually measures when you paste it in, and how to run that check without turning it into an accusation.
AI Detector Checker never identifies which model wrote a text. It returns a 0-100 AI-writing signal score, a score stability range, and one of three bands — Likely human-written, Uncertain or Likely AI-written — for English text of at least 100 words. No result on this site can show that Copilot, rather than any other assistant or a person, produced a passage.
Copilot output inside Microsoft 365 documents
Copilot is usually asked to do one of a small number of workplace jobs, and each leaves a recognisable shape on the page.
- Meeting and thread recaps. A short orientation sentence, then decisions, then owners and dates. Every item is the same grammatical shape, because the model is filling a slot rather than reporting what actually happened in the room.
- Status and project reports. Progress, risks, next steps, in that order, with each risk phrased as a balanced pair: a concern followed by a mitigation. Real risk registers are lopsided; some risks have no mitigation yet.
- Email drafts. A courteous opening line that restates the recipient’s own message, a middle that answers, and a closing offer to help further. The register stays level even when the underlying subject is not.
- Policy and process text. Definitions first, obligations second, exceptions last, with “ensure”, “align”, “leverage” and “stakeholders” doing a lot of work.
- Slide notes and executive summaries. Three to five parallel bullets, each roughly the same length, each ending on a benefit.
None of this is evidence on its own. Plenty of experienced people write status reports exactly this way, because the format is the point. What matters for review is that Copilot output tends to be uniformly competent: the eighth bullet is as polished as the first, the risk nobody understands yet reads exactly like the one that is fully scoped, and no sentence is noticeably rougher than its neighbours. Human workplace writing usually has a tired paragraph somewhere in it.
Checking a colleague’s draft: consent comes first
Workplace documents are the one setting where running a detector is a decision about a person, not about a text. A student submits work for assessment; a colleague sends you a draft to get a job done. Before you paste someone else’s document into any detector, including this one, three things are worth being honest about.
- Whether your organisation permits AI assistance at all. In many teams Copilot is licensed, encouraged and paid for. Checking a document for AI writing in an organisation that has told people to use Copilot is measuring compliance with a rule that does not exist.
- Whether the text is yours to submit. Pasting a colleague’s draft into a third-party website sends their words off your network. Contracts, personnel matters, customer data and anything under NDA should not be pasted into this tool or any other. Our privacy policy sets out what happens to submitted text; the safer rule is not to submit it.
- What you would do with a result. If the answer is “raise it with them”, the conversation is available to you now, without the check. If the answer is “put it in a file”, stop. A score is a review signal, not proof, and it is never enough on its own for a decision about a person’s job.
The defensible use of this page is on your own drafts, on text you have been asked to review with the writer’s knowledge, or on published material where authorship is a factual question rather than a disciplinary one.
What Copilot output looks like, and what the detector measures
The two columns below are deliberately not the same list. The left column is what people notice when they read a Copilot-assisted document. The right column is what the frozen engine behind this site actually computes. They overlap, but they are not the same thing, and confusing them is the main way detector results get over-read.
| What a reader notices in Copilot text | What AI Detector Checker measures |
|---|---|
| Parallel bullets of near-identical length | How closely the whole passage matches patterns a frozen classifier learned from human and AI text across many sources |
| House-style vocabulary: align, ensure, leverage, stakeholders | Variation in sentence length across the passage |
| Every risk paired with a tidy mitigation | Matches against a fixed list of 60 generic framing phrases |
| A level, unbothered register from start to finish | The presence or absence of numbers, named people and places, and first-person reference |
| Formatting that came from the app, not the writer | Nothing about formatting, and nothing about which product produced the text |
| Your knowledge of who wrote it and how they usually write | Nothing at all — the engine sees the pasted words and nothing else |
The right column is why the same score can appear on a Copilot recap and on a carefully written human compliance memo. Both are uniform, both avoid first person, both are short on specifics. The left column is the part only you can supply.
Running the check, and reading what comes back
Paste at least 100 words of English into the tool above, or on the homepage — it is the same detector. Paste the body of the document rather than headings, tables and signature blocks, because those are formatting rather than writing. If a document has several distinct parts, check them separately: the compare mode takes up to three passages and reports each one on its own, without averaging them.
What comes back is a score, a stability range and a band. What an AI detector score means explains each of those in full, and how the detector analyzes text explains where the number comes from. The short version for workplace documents: a Likely AI-written band means the writing matches AI patterns strongly at that length, not that Copilot wrote it; a Likely human-written band means no strong signal was found, which is not a clearance, because in our sealed test a large share of AI text also landed there; Uncertain is where edited and partly drafted text usually lands, and that describes most Copilot documents, since somebody nearly always edits them.
Where this goes wrong
Workplace writing is unusually exposed to false signals. Templates, mandated phrasing, compliance language, and text written in a second language all push a passage towards the patterns the engine associates with AI. Short texts are riskier still: at 100 to 149 words about 2.4 in every 100 human texts drew a wrong AI signal in our sealed test, against about 2 in every 1,000 at 150 words or more. The full figures, including the two release targets the test did not meet, are on the evidence page.
If your own document was flagged, what to do if your text is flagged sets out the sensible next steps. English is the only language this detector has been validated on; language support explains why a result on anything else is unsupported.
What the evidence does and does not cover
There is no Copilot-specific figure on this site, and there will not be one until a sealed test includes Copilot output. The sealed test that released this engine covered two current OpenAI models and one model lineage the engine had never seen in training. Everything on this page is a description of writing patterns, not a measured detection rate for Microsoft 365 Copilot. Anyone quoting a per-product accuracy number for Copilot detection — including us — should be asked for the denominator.
Microsoft Copilot detector FAQ
Can AI Detector Checker tell whether Microsoft 365 Copilot wrote a document?
No. The engine returns a 0-100 AI-writing signal score, a stability range and one of three bands for the whole passage. It never names a product, so it cannot separate Copilot output from another assistant’s output, or from a person who writes tidy status reports.
Should I check a colleague’s draft?
Only with their knowledge, and only if a result would change something you could not settle by asking them. If your organisation licenses Copilot and encourages its use, a check measures compliance with a rule that does not exist. Never paste contracts, personnel matters, customer data or anything under NDA into this or any other detector.
Why do human-written workplace documents get flagged?
Templates, mandated phrasing, compliance language and English written by a second-language writer all push a passage towards the patterns the engine associates with AI. Short documents are riskier still: at 100 to 149 words about 2.4 in every 100 human texts drew a wrong AI signal in our sealed test, against about 2 in every 1,000 at 150 words or more.
What part of a document should I paste?
The body prose, not headings, tables or signature blocks — those are formatting rather than writing. If a document has several distinct parts, check them separately: compare mode takes up to three passages and reports each on its own, never averaged into a single verdict.
Is there a published detection rate for Copilot?
No. The sealed test covered two current OpenAI models and one model lineage the engine had never seen in training; Copilot was not among them. Anyone quoting a per-product accuracy figure for Copilot detection, including us, should be asked for the denominator.