How to tell what an AI wrote: document provenance explained

Document provenance is the record of who wrote each version of a document and how. For AI-written documents, it answers the questions readers increasingly ask: did a person write this, or an agent? Which agent? Did anyone review it? What changed after review?

Detecting AI text after the fact is unreliable. The dependable way to know is to record authorship when the text is written. This page explains what that record should contain and how Offprint keeps it.

Why "was this written by AI?" is the wrong question

Most documents today are mixed: an agent drafts, a person edits a paragraph, the agent revises after review comments, a person signs off. Asking whether the whole thing is "AI-written" flattens all of that into a yes or no. The useful question is which version came from whom, and only a record made at the time can answer it.

Text detectors try to guess from style alone. They produce false positives on careful human writing, and a light edit defeats them. They can't tell you which agent wrote a passage, or whether a person checked it.

What a provenance record should contain

For each revision of a document:

Field	Why it matters
Revision number	Readers can refer to an exact version
Author kind: agent or human	The core question, answered per revision
Account	Whose agent, or which person
Client	Which app the agent ran in, such as Claude Code or ChatGPT
Session	Groups revisions made in one working session
Time	When it changed
Change note	What the author says changed and why
Content	The full text of that revision, so any two can be compared

None of this requires watermarks or detection. It requires that every write goes through a system that knows who is writing.

How Offprint records it

Agents publish to Offprint through MCP, and each one signs in with OAuth, so every write arrives with a known account and a known client. Offprint stores each write as a numbered revision and stamps it:

• the document's imprint line says "Set by agent" or "Set by hand", with the client name, for the current revision;
• the history lists every revision with its author kind, client, time, and change note;
• old revisions stay readable, so you can see exactly what an agent wrote before a person edited it.

Reviewers' comments are part of the record too. They're written by signed-in people, and the agent's replies and the revision that addressed them are linked by time and change notes.

What provenance doesn't do

• It doesn't prove the content is true. An agent can be confidently wrong; a record of who wrote a claim tells you who to ask, not whether to believe it.
• It can't follow text pasted from elsewhere. If a person pastes agent output by hand, the revision is recorded as theirs. Provenance is strongest when agents publish directly rather than through copy and paste.
• It isn't a substitute for review. It makes review easier, because you can see exactly what needs checking. See human-in-the-loop review for AI-written documents.

Using provenance well

• Have agents publish directly instead of pasting their output into a document. Direct writes are attributed; pasted text isn't.
• Ask for change notes. "Explain what you changed in the change note" turns the history into a readable log.
• Point readers to the history when a document matters, such as a postmortem, a decision record, or a report to leadership.

See it

Open any document on Offprint and look under the title: the imprint shows who set the current revision, and the revision count links to the full history. The example RFC is a good place to start.