Why a clean score proves nothing
Detection is asymmetric. A high score is weak evidence of generation. A low score is not evidence of authorship, because most images in the world are neither generated nor made by you.
It is also fragile in the wrong direction for creators. Retouching, upscaling, noise reduction and generative fill used as ordinary parts of a professional workflow all push scores upward, which means the more finished your work is, the more likely it is to be flagged.
What actually evidences authorship
Every item here is difficult to fabricate after an accusation and trivial to produce if you were working normally. That asymmetry is what makes them persuasive.
Strong
- Layered working file with edit history
- Raw camera files with EXIF intact
- Dated intermediate exports
- Content Credentials signed at export
- Screen recordings or timelapse of the work
Weak
- A low detector score
- The finished JPEG on its own
- A verbal account of your process
- Social posts of the final image only
- A colleague vouching for you
Build the evidence as you work
-
Keep the layered file, always
Never flatten and discard. A working document with named layers, adjustment history and the mistakes you corrected is the single most convincing artefact you can produce.
-
Export three dated progress versions
Early, middle and near-final. Thirty seconds of work per project, and it demonstrates development rather than a finished object appearing at once.
-
Keep raw captures with metadata
Camera raw files carry device, lens, exposure and time. Archive them rather than deleting after export, and do not strip metadata on the way through.
-
Turn on Content Credentials
Where your software supports signing at export, use it. A signed manifest is cryptographic rather than statistical, and it is the only evidence in this list that a machine can verify.
-
Record your tool use
If you used generative fill to remove a lamppost, note it. Disclosed assistive editing is normal. Discovered assistive editing looks like concealment.
If a client questions a piece
Answer with material rather than argument. Send the layered file and two progress exports before discussing the score at all. Most challenges end there, because the evidence is obviously not something you assembled that afternoon.
Where the client is relying on a detector result, it is fair to ask which tool, which threshold and on which file. A client checking a compressed copy pulled from a proof gallery is checking the worst version in the chain, and saying so politely resolves a surprising number of these.
Put it in the contract
The cleanest fix is contractual rather than technical. A short clause removes the ambiguity that these disputes grow in.
- State what tools are permitted, including assistive editing, upscaling and generative fill, and whether disclosure is required.
- Agree what evidence settles a question, naming working files rather than detector scores.
- Say who checks and on which file, so nobody is scoring a compressed proof.
- Set a response window, so a challenge does not sit open against your invoice indefinitely.
The reputational side
An accusation does damage before it is resolved, particularly in public. The instinct is to argue in the same place it was made, which almost never helps and usually extends the conversation.
Respond once, with material. Post a short clip of the layered file or the progress versions rather than a paragraph of denial. People who were undecided find that persuasive, and people who were not were never going to be moved by an argument.
Then stop. Continuing to engage keeps the accusation visible and rewards whoever made it with attention. Where a claim is repeated by an account with reach, a direct message with the evidence attached tends to work better than a public reply.
If work or income is genuinely at risk, this becomes a legal question rather than a technical one. A false public accusation of fraud against a professional is actionable in many jurisdictions, and the evidence you kept is what makes that case straightforward.