Why a detector score isn't proof
A percentage is an opinion held by a model. Turnitin, to its credit, says this itself: its score "should not be used as the sole basis for adverse actions against a student," its sentence-level false-positive rate is around 4%, and it suppresses all scores under 20% because it can't trust them. When Washington State University audited its own integrity cases, 33% of AI-misuse hearings ended in "not responsible" findings, and the university cancelled the product.
The deeper problem is structural. A probability can't be cross-examined. When a student says "I wrote this" and a model says "76%," there is no procedure that resolves the disagreement, because there is no underlying fact the two parties can examine together. Proof has to be something both sides can look at.
What's actually held up
Look at how the recent cases were resolved, and a pattern appears: process evidence wins, scores lose.
In February 2026, a New York State Supreme Court judge reversed Adelphi University's discipline of a student flagged by an AI detector and ordered his record expunged, after the evidence showed his work was his own, supported by the university's own tutoring program. The student's attorney called the ruling groundbreaking on exactly the point that matters here: due process requires evidence, and a score isn't evidence.
In Wake County, North Carolina, a freshman was flagged by three detectors at once (62%, 75%, 87%, three tools, three different answers). What cleared her was not a fourth detector. It was a teacher reviewing the document's version history: a record, however coarse, of the writing process. Her district's guidance now says to use "multiple measures, such as reviewing a student's writing process and work history."
And in the Palo Alto case, the family's defense was built from that same coarse record, 1,162 pages of it, and it still left enough ambiguity for the school to hold its position. The lesson cuts both ways: process evidence is what works, and snapshot-grade process evidence is not always enough.
The three signals that matter
Real authorship evidence comes down to three observable things:
1. Keystroke timing. Human typing has a fingerprint. Bursts of fluent text, stalls at hard sentences, a rhythm of small corrections made mid-thought. Plotted on a timeline, real composition looks jagged and alive. Pasted text appears as a single instant; an autotyper produces a flat, metronomic rectangle no human hand makes. This is the difference between inferring authorship from prose style and seeing it in the physical act.
2. Paste events. Where text came from matters more than what it says. A trustworthy record captures every paste, its size, and when it happened, under a policy the teacher sets per assignment: allow pastes, mark them, or block them. A student who drafts honestly has nothing to explain. A document that arrived in three large pastes explains itself.
3. Focus and session history. When did the student start, pause, leave the tab, come back? A record of sessions and attention gaps completes the picture: an essay written across four evening sessions with natural breaks reads very differently from one that materialized in eleven minutes of perfect focus.
None of these signals read the student's prose. They watch what the student did, not how the student sounds. A multilingual writer's keystroke record looks like everyone else's, which is precisely why process evidence avoids the bias problem that plagues detectors.
How Manupropria packages it
Manupropria is a writing environment that records all three signals as the student writes. The student opens your assignment link and just writes; there is nothing to install and nothing to remember to turn on. Behind the canvas, the app builds the record: keystroke-level timing, paste events under your policy, session and focus history, plus tamper detection so the record itself is trustworthy.
On your side, every submission gets a visual timeline you can read in seconds: typing rhythm above the line, revisions below, pastes and session marks flagged. Eleven independent signals summarize whether the process looks human-shaped, with no composite "AI score" anywhere, because we refuse to ship a number that pretends to certainty. When you need to hand evidence to a parent, a counselor, or an administrator, one click exports the proof as a PDF.
The final text is encrypted at the application layer, students retain their work, and we never train AI on student writing. Proof of authorship shouldn't cost the student their privacy.
Start before the question is asked
Every case above shares one feature: the evidence had to be reconstructed after the accusation, when it was already adversarial. The fix is to let the evidence accumulate before any question exists. Assign writing in an environment that records the process, and every student walks out of every assignment already holding the proof.