AI detection, false positives, and keeping your provenance.
Detection tools are now routine and they are not reliable enough to be treated as evidence on their own. That cuts both ways: students who did nothing wrong get flagged, and the only durable protection is being able to show how the work came to exist.
Written by Annelies VosMSN, CNE · keeper of standards · 2026-05-26
What these tools actually measure
They estimate statistical properties of text — how predictable each word is given the ones before it — and report a probability. They do not detect authorship and they cannot know who typed anything. A high score means the writing is unsurprising, which is not the same as it being generated.
That distinction matters because plenty of honest academic writing is unsurprising by design. Clinical documentation, methods sections, technical descriptions and the careful hedged prose that graduate programs actively teach all score higher than personal essays do. Writing in a second language raises scores too, because non-native writers often use more conventional constructions.
Provenance is the real protection
Nothing you can do to your prose reliably lowers a detector's score, and trying to is a bad use of your time. What does protect you is being able to show the work's history, and that costs nothing if you build the habit before you need it.
Version history is the strongest single artifact. Writing in a tool that saves revisions means you can show a document growing over days, with false starts and reordered paragraphs. A file that appeared complete in one save looks the same whether it was generated or pasted in from elsewhere, and you will not be able to tell the difference either.
- ✦Draft in a tool with version history, and leave it turned on
- ✦Keep your notes, outlines and the sources you read, dated
- ✦Save marked-up PDFs and highlighted readings rather than deleting them
- ✦Keep the messy intermediate drafts instead of overwriting one file
- ✦Where a tool helped, record which and for what, before anyone asks
If you are wrongly accused
Ask what the specific allegation is and what evidence supports it, in writing. If the answer is a detector score alone, say so plainly and calmly: these tools report probabilities, they produce false positives at rates their own vendors acknowledge, and a score is not a finding.
Then offer your provenance. Version history, notes, drafts, sources. Offer to discuss the content of the work, since you wrote it and can explain the choices in it, which is something no generated submission can survive. Follow your institution's process rather than arguing informally, and find out whether you are entitled to bring someone with you, because in many places you are.
Using these tools honestly
Most institutions now allow some assistance and require you to say so, and the policies differ enough that yours is worth reading properly rather than assuming. Commonly permitted: brainstorming, explaining a concept you did not follow, checking grammar, suggesting structure. Commonly prohibited: generating text you submit as your own writing.
The practical rule that keeps people out of trouble is that you should be able to explain and defend every sentence in your submission as your own reasoning. If a paragraph contains a claim you could not justify to your instructor, it does not belong in your paper regardless of where it came from.
The thing this dispatch will not tell you
There is a large amount of advice online about defeating detectors — rewording tools, deliberate errors, paraphrasing services. None of it is covered here and none of it is recommended, for a straightforward reason: it addresses the appearance of a problem rather than the problem, and it produces worse writing.
The version that actually holds is unglamorous. Do the work, keep the trail, and be able to talk about what you wrote. If the reason you are looking for shortcuts is that the volume is genuinely unmanageable, that is a real problem with an honest answer — help sized to what is actually failing — rather than one to solve by making generated text harder to spot.
Questions this raises.
Can a detector prove I used AI?
No. It reports a statistical estimate, not a finding of authorship, and vendors acknowledge false positives. Most institutions' own guidance says a score should prompt a conversation rather than settle one. If a process is treating a percentage as proof, that is worth raising directly and in writing.
Why does my own writing get flagged?
Usually because it is conventional, which academic writing is trained to be. Methods sections, clinical documentation and formal hedged prose all score higher than casual writing. Writing in a second language raises scores as well, since non-native writers tend toward more predictable constructions. None of that indicates anything about how the text was produced.
What is the single best habit to protect myself?
Draft in something that keeps version history and never turn it off. A document you can show growing across several days, with abandoned paragraphs and reordering visible, is the most persuasive evidence available. It costs nothing, it is invisible until you need it, and it is far more useful than anything you could do to the prose itself.
Annelies Vos
MSN, CNE · keeper of standards. Writes for the dispatch on what actually moves marks in this subject, and coaches the students who bring it to the guild. The rest of the guild.
Where to go from here.
Tell the guild your tale.
Name the course, the program and the date it is due. A mentor credentialed in that subject reads it and replies with counsel and a figure within two hours. The reading costs nothing, and sometimes the answer is an honest no.
SEND A RAVEN