Digital publisher · News and features, thirty staff
Publishing AI images with provenance that survives the CMS
The image desk was receiving generated art and stock alternates several times a day. Editors resized, cropped and re-exported, which stripped the provenance metadata the desk had been relying on to answer the only question that mattered: where did this come from?
- 3x
- Re-exports before a typical hero image publishes
- 0
- Manifests surviving the old export path
- Image desk
- Where the gate now sits
Four done, one to go.
The problem was structural, not editorial
Provenance lives in the file. Every resize and re-encode is a chance to lose it, and the standard export path was losing it consistently, so the desk had no record for any image older than a week.
Editors were not being careless. The tooling discarded metadata by default, and nothing in the workflow noticed because a stripped file still renders perfectly.
What the desk checks now
Every incoming image is inspected before it is placed: manifest or Content Credentials present, generator recorded where it exists, dimensions and compression sane for the slot, and no frame that is effectively blank.
- Keep the original file; export from a copy so the source keeps its metadata.
- Record the generator and the prompt author in the story record, not only in the file.
- Check alt text as part of the same pass, since it is written once and corrected never.
- Flag anything involving a recognisable person for a human decision.
The limit worth stating
No check decides whether an image is appropriate to publish, and none proves a likeness is licensed. The desk's rule is that the automated pass answers what the file is and the editor answers what it is for.
Other engagements
Design and build, eight people
Signing off AI-assisted client work without hiring QA
Series A, fourteen engineers
Reviewing agent-written pull requests without stalling the queue
Client names withheld. Figures describe the pattern of these engagements, not audited outcomes.
Your own file
The first report is the useful one.
Run a single deliverable through the console and compare its findings with what your current review already catches.