What is writing provenance? A plain-language definition

The definition

Writing provenance is a verifiable record of how a piece of writing came to be. Not a judgment about the finished text — a record of the process: when the document grew, how it changed, what was typed and what arrived some other way, captured while the writing happens (with the writer's consent), sealed cryptographically when the writer finishes, and checkable afterward by anyone without trusting the writer, their institution, or the software vendor.

The word "provenance" is borrowed deliberately. In the art world, provenance is the documented chain of custody that tells you a painting is what it claims to be — not an expert squinting at brushstrokes and guessing, but a paper trail. In data science, the W3C's PROV standard defines provenance as the record of the entities and activities involved in producing a thing, used to assess its reliability. Writing provenance applies the same idea to documents: the question "did a person write this?" gets answered by a record, not an estimate.

Provenance vs. detection: the one distinction that matters

Everything about AI detection and its failures comes down to a single structural fact: a detector only ever sees the finished text. From word choices and sentence rhythms it computes a probability that the text is machine-generated. That is an inference — and inference about polished prose is unreliable in both directions. Turnitin, the most widely deployed detector in education, publicly cautions that its score should not be the sole basis for an integrity decision; Vanderbilt University disabled the detector entirely after doing the false-positive arithmetic for its own campus, and it is one of a growing, verified list of universities that reached the same conclusion.

Provenance inverts the direction of proof:

	AI detection	Writing provenance
Looks at	The finished text	The writing process
Produces	A probability	A sealed record
Created	After submission, by the accuser's tool	During writing, with the writer's consent
Fails by	False positives and false negatives	Absence (no record = no claim, nothing more)
Serves	The person checking	The person who did the work

The last two rows are the ones people miss. A detector's failure mode is an error that lands on someone — most painfully on the honest writer it misreads. A provenance record's failure mode is simply not existing, which proves nothing either way. And detection is structurally adversarial — a tool pointed at writers — while provenance is evidence a writer carries for themselves. It is the difference between being scanned and holding a receipt.

What makes a provenance record trustworthy

A revision history, a draft folder, or a browser trail is already a kind of process evidence — that is why integrity panels accept them, as our evidence hierarchy explains. What promotes a process record from "helpful context" to "verifiable proof" is four properties:

1. Tamper-evident. The record is cryptographically sealed when the writer finishes. Change anything afterward — the text, the timestamps, the record itself — and the seal visibly breaks. This is what ordinary version history lacks: it can be simulated or replayed with free public tools, and nothing in it breaks when it is.
2. Consent-based. Recording is off until the writer turns it on for a specific document. This is not just an ethical preference; it is what makes the record attributable. Evidence a person knowingly created and attests to carries weight that ambient surveillance never can — and a system that records everyone all the time is a workplace-monitoring tool, not an evidence tool.
3. Attributable. A named person stands behind the record. "Account X edited this document" is weak; "this person attests, under their own identity, that this sealed record describes their work" is an evidentiary statement someone can be held to.
4. Independently verifiable. Anyone — an instructor, an integrity panel, an editor, a court — can check the seal without installing software and without trusting the vendor who made it. A proof you can only verify inside one company's platform is a subscription, not a proof.

If a product claims "provenance" and is missing one of these — typically the tamper evidence or the independent verification — it is a logging feature wearing the word.

Isn't this what C2PA already does?

Related, and worth knowing about. C2PA (the Coalition for Content Provenance and Authenticity, whose members include Adobe, Microsoft, and camera manufacturers) defines a standard for attaching signed provenance metadata — "Content Credentials" — to media files. It answers where did this file come from and what was done to it, and it is excellent for images, video, and audio, where the capture device can sign at the moment of creation.

Writing is the hard case the media standard doesn't reach. A document is not captured in a moment; it is accumulated over hours and days of edits, across sessions, with pastes and rewrites and deletions. Signing the finished file tells you who exported it, not who wrote it. Writing provenance is the layer that records the accumulation itself — and it can then use standards like C2PA as the envelope that carries the sealed record with the final document. The two are complementary: C2PA is the tamper-evident container; the writing-process record is what makes the container worth opening.

What a provenance record does and does not claim

Honesty about limits is part of the definition, so:

• It does show the document's growth over time at fine granularity, what was typed versus pasted, and that the record has not been altered since sealing.
• It does not read minds. A record shows the process; it cannot show intent, and no system can prove the absence of every conceivable assistance. What it can make is the strongest available positive case: this person did this work, here is the sealed record, check it yourself.
• It proves nothing by its absence. A writer without a provenance record is exactly where every writer was before — entitled to the presumption of innocence and to the ordinary layers of evidence.

Who actually needs this

• Students, for whom a false flag can mean a failed course, a transcript notation, or — for international students — a visa problem. The first-48-hours guide exists because they usually meet this problem after the accusation; provenance is the before-the-accusation answer.
• Educators and integrity officers, who need something sturdier than a detector score to act on and fairer than an interrogation to offer. (We write for them separately in our educators section.)
• Professionals — journalists, researchers, applicants, freelancers — anyone whose work product is now routinely run through a detector by a client or editor who will never convene a hearing about it.

How Authmarks implements it

Authmarks is a writing-provenance system built on exactly the four properties above: recording is off by default and enabled per document with explicit consent; the process record is hash-chained and sealed when you finish; the sealed record belongs to you and travels with the document; and anyone can verify it in a browser, free, without an account. It works where you already write — desktop app, browser extension, and LMS integration — so the record accumulates while you work, not as an extra chore after.

The pitch in one sentence: detectors guess about your finished text, and you absorb their mistakes — provenance lets your writing carry its own proof.