Proving you wrote it: authorship evidence that holds up

Why the finished text is weak evidence

Any judgement made from the finished text alone, whether by a detector or by a reader's instinct, is an inference from patterns, and patterns can mislead in both directions. The guide to what a detector score means explains why a high score is not proof. Research on authorship verification, the long-running PAN shared tasks among them, shows how hard it is to establish who wrote a text from style alone, even for specialist systems with samples of a known author's writing.

Evidence that carries weight

What does not help

• Running your text through other detectors until one says human. Detectors disagree for reasons covered in why AI detectors disagree, so a low score elsewhere proves little.
• Rewriting the text to lower its score. It changes the product, not the history, and can look worse.
• Deleting drafts or history to tidy up. The mess is the evidence.

Worked example: what a version history looks like

The counts below are invented but realistic. A 1,800-word essay written over nine days in an editor that records every save. The history is the evidence, and the shape of it is what a reader checks.

No detector can produce this table, and no paraphrase can fake it after the fact. The paste on day 6 is worth noticing: an honest history contains pastes, and a reviewer who sees one should look at where it came from before drawing a conclusion. Version history is checkable evidence of process, which is why the guide to what a detector score means ends where this guide begins.

Common mistakes writers make

• Writing in a scratch file and pasting the finished text into the editor that keeps history, which produces exactly the one-save history that looks worst.
• Deleting the notes file after submission to tidy up. Keep it with the project for as long as the work can be questioned.
• Relying on document properties such as author name and editing time. They are trivially changed and are often stripped when a file is uploaded.
• Responding to a flag with another detector's low score, which invites a third detector and settles nothing.
• Disclosing tool use vaguely. A note that says which tool, on which passages, for what purpose is far stronger than a general statement.

Expert tips

• Turn on version history before you start, in whatever editor you use, and write in that file from the first note to the last edit.
• Save a dated copy at the end of each session with the date in the file name; a folder of nine drafts is evidence a reader can open in a minute.
• Keep the reading list with page numbers and the date you read each source, since it ties the argument to the research behind it.
• If you use any tool, log it as you go: which tool, which passage, what it did. A contemporaneous note is stronger than a recollection.
• When asked to respond, assemble one packet: the history export, the drafts, the notes, the tool log and a one-page statement of how the work was done, and offer to walk through it in person.

Assembled this way, a response packet takes an afternoon and answers the question that matters, which is how the text came to exist. Institutions that ask for it get better evidence than any detector can supply, and writers who keep it never have to reconstruct a month of work from memory.

Metadata and content credentials for text

Two kinds of technical provenance exist for files, and both are weaker for text than writers hope. Document metadata (author, created and modified dates, total editing time) is set by the editor and can be changed by anyone; more to the point, most submission systems, email gateways and web uploads strip or rewrite it, so it rarely arrives intact even when it is honest. The guide to metadata survival testing shows how to find out what a given route preserves. Content credentials, the signed provenance manifests that the C2PA specification defines for images and video, are the stronger form: they record which tool created the content and what was done to it, and a broken signature is detectable. They only help a writer if the editor they use signs the file and the platform they submit to keeps the manifest, which is what verifying content credentials and testing content credentials pipelines check. For most text today, neither exists end to end, which leaves version history and drafts as the evidence that reliably survives.

Honest use of writing tools

Many writers use grammar checkers, translation aids or drafting assistants for part of their work, and many institutions and publishers allow some of that. The simplest protection is to know the rules where you submit, keep a note of what tools you used and for what, and disclose it when asked. Provenance standards such as C2PA content credentials show where media authenticity is heading, and text watermarks may mark some machine output, but for most writing today your own record of the process remains the most reliable evidence of authorship.

Institutions can help by saying in advance what evidence they will look at. Weber-Wulff and colleagues, after testing fourteen detection tools, concluded that detector output should not be used as the basis for accusations; process evidence is the fairer ground for everyone.

This guide is part of the testing AI systems hub. It is best read alongside what an ai detector score means for a writer, teacher or editor and text watermarking for language models, which cover the neighbouring questions.