We ran an open AI detector over five passages published before machine generation existed. It scored a club notice from 1923 at 97.5% machine-generated. Nobody in 1923 had a language model. The score is wrong by construction, and it is wrong at near-certainty.
The measurement
| Published | Scored | Verdict |
|---|---|---|
| 1923 | 97.5% | flagged as machine-generated |
| 1877 | 22.1% | not flagged |
| 1912 | 21.8% | not flagged |
| 1890 | 15.4% | not flagged |
| 1904 | 5.8% | not flagged |
Model: onnx-community/tmr-ai-text-detector-ONNX, q8 quantisation,
measured 2026-08-07. 1 of 5
passages flagged. Reproduce with node sites/draftrecord/tools/measure-detector.mjs.
What we did not find
This page was built expecting a second result, and did not get it. The plan was to reorder each passage's sentences and show the score moving — the same words, in a different order, scored differently. It did not move. The largest swing across the corpus was 0.13 of a percentage point.
So that argument is not made here. On this evidence the model was consistent with itself; what it was not is correct. We are telling you about the experiment that failed because a page that only reports its successes is the kind of page this site exists to warn you about.
What this means if you have been accused
- A detector score is one model's guess. This one guessed 97% about a document from 1923.
- It is not the model your institution used, and you cannot inspect the one that was.
- A published false-positive rate of one per cent sounds small until it is your document, and it is not distributed evenly: writers whose first language is not English are flagged at several times the rate of native speakers.
- What is checkable is your file's own record of how it was written. That is what the tool on this site reads.
Run it on your own text
You can score your own writing with the same model, in this page. It is a 120 MB download the first time — the model itself, fetched from Hugging Face — and nothing is sent anywhere: the text you paste is scored on your own device and never leaves it. Your browser will keep the model afterwards.
0 words
The download happens once. Nothing you paste is transmitted.
What this cannot tell you
- Word does not follow the ECMA specification for revision save identifiers, so the exact order of editing sessions cannot be recovered from the file.
- A writing history shows that editing happened. It does not show who did the writing.
- No writing history in a file is not a sign of wrongdoing. Exports, converters and generators all produce documents with none.
- This is not a detection result. It does not answer whether any part of the document was written with AI assistance, and it cannot.