A student emails her professor. The paper she spent three weeks on came back with a comment: "This reads as AI-generated. Please see me." She did not use AI. She has nothing to show for the three weeks except the paper itself — the finished thing, with no trace of how it got made. She is now in the strange position of having to prove a negative.
This is happening constantly, and it is going to get worse.
Detectors don't detect what they claim to
AI-detection tools promise to tell you whether a piece of text was written by a human or a machine. They cannot reliably do this, and the people who build them mostly admit it in the fine print.
The reason is structural. These tools look for statistical signatures — low "perplexity," even sentence rhythm, predictable word choice. The trouble is that good academic writing has exactly those properties. Clear, conventional, well-edited prose looks, statistically, a lot like what a language model produces. So does writing by non-native English speakers, who often learned in a more formal register. So does anything that has been through a grammar checker.
The result is false positives, and they are not rare. Multiple studies have shown detectors flagging human-written text — including, famously, the US Constitution — as AI-generated. A tool that confidently mislabels real writing is not a detector. It is a random number generator with a trust problem.
The burden of proof quietly flipped
Here is the part that should bother everyone more than it does.
For a long time, the assumption in academic work was innocence. If someone accused you of cheating, they had to show it. Now, in a lot of classrooms, a detector score is treated as evidence, and the student is asked to demonstrate that they wrote their own work.
Think about how hard that is. How do you prove you wrote something? "Trust me" is not proof. A few saved drafts help, but most people overwrite the same file. Your memory of writing it is not admissible. The finished document — the only artifact most writers keep — says nothing about its own origin.
We built an entire academic culture around the final product and threw away the process. Now the process is the only thing that would actually clear you, and almost nobody has it.
A better detector won't fix this
The instinct is to want a more accurate detector. Catch the real cheaters, clear the honest students. But you cannot build your way out of this with classification. Models that generate text and models that detect it are locked in an arms race the detectors are losing, and every percentage point of accuracy comes with false positives that land on real people.
The honest move is to stop asking "was this written by AI?" — a question we cannot answer — and start asking "can this person show how the work came together?" — a question that has an answer.
Keep the receipts
The thing that actually protects you is a record of your process. Not a vibe, not a promise. Evidence:
- A version history that shows the document growing over time, the way real writing does — in fits, with deletions, with a paragraph that moves three times before it settles.
- A clear trail of the sources you read and cited, so your argument is anchored to real references rather than fabricated ones.
- An honest log of where you did use AI — for proofreading, for finding a source, for asking a question about your own notes — because using AI as an assistant is fine, and pretending you didn't is what gets people in trouble.
If you can hand someone that, the detector score stops mattering. You are not arguing about statistics. You are showing your work.
This is, bluntly, why we built Folio the way we did. The assistant only answers from your own sources, so it can't put words in your mouth. And when you're done, you can export an authorship report — a timeline of how the document evolved, the sources you cited, and exactly where AI assisted. It is not a claim of innocence. It is a receipt.
You shouldn't have to prove you wrote your own essay. But until the culture catches up with the tools, the people who keep the receipts are the ones who will be fine.
Folio gives you an authorship and proof-of-work report for every document — so "I wrote this" comes with evidence. Join the waitlist.