Let a student prove authorship
A student keeps a record of their drafts as they write. If they are ever questioned, that record is checkable by you, by them, and by anyone else. It is the evidence authorship has always rested on.
You have to make decisions that are fair, that a student can appeal, and that hold up when a parent or a lawyer asks how you reached them. Here is what we can and cannot give you.
If you are looking for a tool that tells you which essays were written by AI, no such tool exists, including ours. Anyone who tells you otherwise is selling you a number they cannot stand behind.
The reason matters. A style-based detector has to be wrong sometimes, and when it is wrong it is wrong about a real student. A 2023 study in Patterns (Liang et al.) measured a 61.3% average false-positive rate across seven detectors of that period, on 91 TOEFL essays by non-native English writers; all seven unanimously misclassified 19.8% of them. That finding is contested: a peer-reviewed 2026 EACL study retesting three detector families, in Czech, found no systematic bias. We think the honest reading is that this is a per-tool, per-population question to be measured, not a settled fact either way, and that an institution should ask for the rate measured on writing like its own students’.
What is not contested is the arithmetic. Vanderbilt University disabled its AI detector in August 2023, saying it did not believe AI detection software was an effective tool that should be used. In that same published guidance the university did the sum itself: at the vendor’s advertised one percent false-positive rate, against roughly 75,000 papers submitted in 2022, about 750 could have been wrongly flagged in a single year. That is their arithmetic using the vendor’s own claimed figure, not a number we derived.
So we do something narrower, and we can actually defend it.
A student keeps a record of their drafts as they write. If they are ever questioned, that record is checkable by you, by them, and by anyone else. It is the evidence authorship has always rested on.
If a submission arrives with a signed record of how it was made, we read it and verify the signature. For ordinary text files there is usually nothing to read, and we say so rather than guess.
Every result comes with the evidence behind it and a certificate that can be re-checked independently. A decision you can show your working for.
We serve both sides on purpose. The same tool that helps you assess a submission helps a student defend one. If a product only ever works against the student, it is not doing justice, it is doing convenience.
Nothing is added at the end, and nothing is asked of the student beyond writing. The record accumulates while the work does.
In Word or Google Docs, on Windows, Mac or the web. No new tool to learn and no change of habit.
An add-in takes a 64-character fingerprint of the text on the student’s own machine. The words are never sent anywhere, and a fingerprint cannot be turned back into text.
SharePoint, Google Drive or an internal store, with the record tied to the student’s identity in your own directory.
One certificate, re-checkable by anyone, years later. It shows what the record holds, including where there are gaps.
The chain and the certificate this produces are already working: a finished 240-page manuscript has been sealed this way, 26 drafts between April and October 2026, and the certificate verifies against our published key — a check anyone can repeat without trusting us. We provide the plugin through your institution rather than as something a student downloads. We work with your IT to put it in front of the cohorts you choose, alongside the engine, which is also what lets each record be tied to your own directory identities and your own storage.
Most universities cannot send unpublished student work to an outside company, and they should not have to. Our verification engine holds no database, keeps no state, and writes nothing to disk. That makes it straightforward to run inside your own environment, where student work never leaves your control.
SharePoint, Google Drive or an internal store. We provide the connectors; you keep the files and the retention policy.
If you run a commercial AI detector, plug it in with your own key and your own data terms. We report its score as one capped input, weighted by its measured error rate rather than its marketing claim. If you run none, the record still works.
The engine needs no outbound network to build or verify a record. An isolated install is supported.
Signing inside your estate proves integrity and order: alter one character, reorder two drafts or remove one, and the chain breaks. What a signature of your own cannot establish is when a record was made, because the dates are supplied by whoever built the chain — and that is as true of an institution as of a student. For that, a record needs a timestamp from someone with no stake in the outcome. We offer that as a counter-signature over a single 64-character hash, which carries no student work and no identity. An air-gapped deployment gets integrity evidence, not dated evidence, and we will put that in writing rather than let it be assumed.
The engine and its document endpoint are built and tested, and they run today. They are not yet packaged as an on-premise product you can install yourself, and there is no hosted service to log into. An on-premise deployment is a conversation with us during the private beta, not a download. We would rather tell you that now than after a procurement process.
The endpoint takes one document at a time and accepts PDF, plain text, Markdown and HTML, so processing a cohort is a matter of running it across a folder. Theses, coursework, technical reports and engineering submissions are all just documents to it.
What you would get back for each one is a verdict, the evidence, and a certificate. What you would not get is a ranked list of suspects, because that is the artefact that causes the harm. If a department wants that, we are the wrong supplier and we will say so in the first meeting.
We do not. Plagiarism detection asks whether a submission resembles work that already exists, which requires a large licensed library of documents to compare against. We hold no such library and we produce no similarity percentage. The two tools answer different questions, and you will probably keep both.