How to prove your art isn't AI
Detection tools cannot do it: the false-positive research is damning. What works is process evidence, and it has to be collected while you work, not afterwards.
You cannot prove it after the fact by running your own work through a detector. Start there, because it is the thing most people try first and it is worse than useless.
Why detectors cannot clear you
The most-cited evidence is a 2023 Stanford HAI study that ran seven popular AI-text detectors against TOEFL essays written by non-native English speakers. The detectors flagged that human writing as AI-generated 61% of the time on average. One detector flagged 97% of the essays. On essays by native English speakers, the same tools had false-positive rates under 10%.
The mechanism is not mysterious. Detectors look for low variance in word choice and sentence structure. Writers working in a second language (and writers with plain, disciplined, edited prose) produce exactly that pattern. The tool is measuring “predictable writing,” and calling it “machine writing.”
OpenAI withdrew its own AI-text classifier in 2023, citing low accuracy. The company with the most access to how the models work could not build a reliable detector for their output.
A tool that clears you today may flag you tomorrow after a model update. Evidence you cannot rely on twice is not evidence.
What actually works: process, captured as you go
The one thing generated work does not have is a history. That is your asset, and it only exists if you preserve it.
Version history with real timestamps. Not one saved file, the trail. Google Docs and Notion keep revision history automatically. Writers using local files should keep drafts in version control or at minimum dated copies. The signal is not any single version; it is the shape of the edits over time.
Work-in-progress captures. For visual work, this is the strongest evidence available. Layered source files (PSD, .procreate, .blend, .ai) contain the construction of the image. Timelapse recording (built into Procreate and Clip Studio, available as plugins elsewhere) produces a continuous record that is currently very hard to fake convincingly.
Source material. Reference photos with EXIF data. Life-drawing sheets. Recorded interviews behind an article. The raw inputs you worked from.
Sketches, dead ends and rejected versions. Nobody generates the four attempts they discarded. Keep them.
Dated public posting. Posting progress publicly as you go creates third-party timestamps you do not control, which is precisely what makes them worth something.
Content Credentials
The technical route is C2PA / Content Credentials, an open provenance standard backed by a coalition of camera manufacturers, Adobe and others. It attaches cryptographically signed metadata recording how a file was made and edited.
Two honest caveats. It only proves what a signing tool asserted, so it is only as trustworthy as the software in the chain. And metadata is routinely stripped by platforms on upload, so the credential often does not survive the journey to where your audience sees the work.
It is a good standard worth adopting. It is not yet a general answer.
Declaration plus consequence
For most working artists the practical route is not proof but attestation with something at stake: a clear public statement of your process, specific enough to be falsifiable, backed by evidence you could produce on request.
This is exactly how the Authors Guild’s Human Authored certification works: a declaration against published criteria, with a mark that can be revoked. No classifier is involved anywhere in the process.
That is not a weakness in the design. It is the recognition that detection does not work, and that the alternative, a reputation you can lose, has been how professional trust operated for a very long time.