No AI label on a picture? That doesn't make it a photograph.

Image Evidence published its study on October 12 and updated the figures on October 15. It examined 77 AI-generated image files uploaded to Wikimedia Commons, using its own checker. The pages identified their generators. The researchers selected the files by hand, limited each uploader's contribution per generator and excluded identifiable real people.

The checker found a readable provenance signal in 19 of those 77 files. Fourteen carried C2PA Content Credentials, a form of signed information about a file's origin and editing history. Across the six processing steps tested—including conversion, resizing, cropping and a Chrome screenshot—none of those embedded C2PA records survived in the processed outputs. One resizing test covered 13 rather than 14 files because the tool could not write WebP.

Those are results from this sample and these tools, not a failure rate for all AI images. The files were Commons uploads, not fresh downloads from every generator. The checker could not read Google's SynthID or other pixel watermarks. Photoshop, phone galleries and social platforms were not tested. The researchers also disclosed a checker bug fixed before the updated figures. This is a reason to ask what a check examined, rather than treating its answer as universal.

Seven of the study's 14 Content Credentials did not validate. The reported reasons concerned expired or invalid signing certificates; the checker reported no content mismatch in those seven cases. That does not establish the images were genuine photographs—they were selected as AI images. It shows why the word ‘invalid’ needs an explanation before anyone turns it into an accusation of tampering.

A missing record tells you even less. You may have a copy that lost information, an image that never had it, or a file whose signals this particular checker cannot read. None of those possibilities establishes whether the scene happened.

The C2PA's own FAQ describes ways to rediscover separated credentials using fingerprints or invisible watermarks, known as soft bindings. So losing embedded metadata does not prove every possible route back to the history is gone. It does mean a simple empty metadata result should not be sold as an authenticity verdict.

Adobe's content-transparency guidance, updated October 8, describes automatic C2PA metadata for supported generative-AI creation and editing. It also makes a distinction a busy publisher could easily miss: machine-readable metadata does not change the picture's appearance. A visible watermark or caption serves a different purpose.

Adobe says preservation outside its applications depends on external support. That is compatible with the study's warning: attaching a record at creation and retaining a useful record through a particular publishing route are separate questions. The study did not test Adobe's whole system, and Adobe's documentation is not independent evidence that every shared copy retains its history.

For a small team, the useful question is concrete: can the person receiving the final file still inspect the information you intended to send? A successful check on the designer's original says nothing by itself about a screenshot pasted into tomorrow's newsletter.

If the picture makes a consequential claim, ask for the original file or first publication link. Check who published it, when, and what they say it shows. A signed record can help establish a file's history; you still have to examine the caption and the claim attached to it.

For an image your team makes with creative AI tools, keep the original export and a short, plain account of how it was made. Inspect both that export and a copy downloaded from the actual destination. Adobe points users to its Content Authenticity Inspect page for credential checks. Don't upload confidential or unpublished client material to an external checker without permission.

Use a visible disclosure when readers need to know an illustration was generated, rather than assuming hidden information will do that job. That is an editorial recommendation, not a claim that one caption meets every legal requirement.

And when a forwarded picture has no useful history, you can leave the question open. ‘I couldn't establish where this came from’ is a better message to pass on than an unsupported verdict about whether it is real.