Reading the AI prompt record left in a PNG

A PNG file has places for text as well as the picture. Many AI tools note how an image was made there, so a PNG as first received is often decisive evidence on its own.

What gets recorded

tEXt and iTXt chunks store the prompt, negative prompt, seed, sampler, model name and node workflow. Some services also write an AI-generated digital source type into XMP. Origma reads these fields and records which keys were present in the ledger.

How strong the evidence is

A generating tool's work record is hard to produce by accident, so it is strong evidence. It is still plain text anyone could write, so if it conflicts with the pixel detectors the conflict is shown as well.

When it gets erased

Conversion to JPEG or WEBP, messenger and social uploads, image optimisers and screenshots mostly remove these fields. Erased copies are judged by the detectors, so the same picture can rest on different evidence.

When checking

Do not re-save the PNG; scan it as received. If record keys appear in the ledger, that is strong evidence the file came from a generating tool; if not, read the detector result as a range.

Frequently asked

Is the full prompt made public?
We keep which records existed, but the full text is shown only as far as needed, and the same rule applies when handing over a report.
What about a real photo with a fake record added?
If the record and the pixels disagree, it is shown as a conflict and no firm call is made.
Do JPEGs never carry such records?
Some tools write them into EXIF or XMP, but it is rarer than in PNG and easily removed.
And WEBP?
WEBP can hold metadata and is read the same way, but conversion has often removed it.

Read next

This is general guidance, not legal advice. A scan result is a probabilistic assessment, not proof.

Try a file