Have you been falsely accused of using AI?
An academic misconduct allegation over work that is legitimately yours is unnerving — but your school has written rules for this exact situation, and the drafts and version histories you've been stressing over are evidence. Gathered Work puts the policy on one page per institution, quoted word for word from the official sources, dated, and linked to the originals — plus a way to organize your work into a clear packet.
No detector vendor or law firm is behind this site, and it determines nothing about any case — it points to the sources and keeps the record of what they said, and when.
The index, then the records
The institutions. The primary work of this site: one page per institution, quoting its published academic-integrity sources — the policy, the resolution procedure, the appeal path — with retrieval dates, direct links, and a change log. These are potentially relevant official sources, not a determination of which governs any case; confirm applicability with the institution. 176 schools so far (and more added daily), including Western Governors University, Adelphi University, Texas A&M University, University of Maryland, College Park, and San Francisco State University — the index.
The records. For writing already questioned: a free checklist for saving what was received, preserving drafts and version history without modifying originals, and arranging the records chronologically. Alongside it, the eight packet documents — cover sheet, exhibit index, chronology, caption sheets, source list, redaction checklist, transmittal note, and assembly instructions — and a browser-based assembler that stacks finished exhibits into one paginated PDF. All of it is free, and none of it selects or pre-fills anything based on anyone's institution or circumstances. If an outcome could affect enrollment, immigration status, a degree, professional licensing, or accommodations, consider obtaining qualified assistance before responding.
A free local hashing utility lives on the vault page for writers who keep fingerprint logs. It shows that particular files existed at particular times; it does not verify authorship or authentic development, and it is a supporting tool here, not the product.
This site is independent: no detector vendor and no law firm is behind it, and it sells neither detection nor legal services.
The measured performance of the accuser
Detection vendors advertise document-level false-positive rates near 1%. Turnitin's own published figure carries two qualifications: the under-1% document rate applies only to documents already scored at 20% or more AI writing, and the sentence-level rate — the one that governs any single highlighted passage — is stated at around 4%. Independent testing returns higher figures, concentrated on particular writers. Liang et al. (Patterns, 2023) tested seven commercial detectors on 91 human-written TOEFL essays from a Chinese writing forum: an average of 61.3% were flagged as machine-generated, against substantially lower rates on the study's US eighth-grade comparison set. Weber-Wulff et al. (International Journal for Educational Integrity, 2023) tested fourteen tools and found the risk of a false accusation ranging from 0% for one tool to 50% for another, with machine-translated text flagged far more often than text written directly in English.
Who the errors land on is itself a finding. Stowe et al. (preprint, 2025, revised 2026) evaluated sixteen detection systems against student essay datasets and reported that essays by English-language learners were more often classified as machine-generated, the disparity widest for non-White ELL writers; human readers annotating the same essays showed no comparable bias. A 2024 national survey of teens and parents by Common Sense Media found 20% of Black teens reported a teacher had wrongly flagged their schoolwork as AI-generated, against 10% of Latino teens and 7% of white teens — a K–12 finding, cited here at its own scope.
A 2026 preprint tested five keystroke-timing classifiers and reported forgery evasion above 99.8% under its specified attacks, alongside a proof that copy-typing machine text cannot be identified from timing data alone; it did not evaluate every process-tracking design.
Share of human-written text wrongly classified as machine-generated in the cited study. The 61.3% figure is from 2023; no independent replication at comparable scale was located as of September 2026. Sources in the footer.
These studies measure particular detectors on particular datasets; they do not yield one aggregate error rate. What they show is that false positives are real, that measured rates vary widely with the writer and the dataset, and that the writers most affected in the research are often those least positioned to contest a flag. Each study, what it tested, and what it reported: the research register, which is kept current as new work is published.
What each method can support
No method proves authorship, in either direction. That limitation is symmetric. What weight any of these methods receives depends on the institution and its procedures, which is why the place to start is the institution's own published sources. What a writer can document is process: that particular drafts existed in sequence, on dates.
| Method | Asserted claim | Supportable claim |
|---|---|---|
| Detector score | "This text is likely machine-generated" | A statistical estimate with documented, uneven error rates |
| Keystroke tracking | "A person typed this, here" | Timing data, forgeable, carrying no identity information |
| Version history | "This document developed over time" | Process evidence, subject to platform custody and retention |
| Cryptographic timestamp | "This exact draft existed at this time" | Independently verifiable by any party, with no vendor to trust, for as long as file and proof are kept |
A single timestamp establishes little, and a sequence of them documents exactly one thing: that particular files existed over time. It does not establish who created them, how they were created, or whether their development was authentic — the vault page states these limits plainly. What it gives the honest writer is narrower and real: a dated record that a score can be weighed against, verifiable by the party judging it.
Who keeps this kind of record
Students facing integrity proceedings where the score is the case. Graduate students with flagged thesis chapters. Writers whose language background, dialect, or neurotype places them on the wrong side of the measured error distributions. And the professionals who advise them — education attorneys, ombuds, writing-center staff — for whom an organized, verifiable record shortens every case.
The underlying instrument is a century old. Laboratory notebooks established priority and integrity with dated, witnessed, sequential entries long before anyone thought to ask a machine its opinion of a finished page. This is that instrument, applied to writing.