Some image files arrive carrying a signed statement about how they were made — a growing number of editors and AI tools write one. When you upload a photo whose file states it was produced by an AI model, that statement is what lets us offer you the visible AI label on export, even though we did not generate the image ourselves.
Until now we took that statement at face value. We read what the file said and never checked whether the signature covering it actually held up. A file whose statement had been damaged, or written by anyone at all, carried exactly as much weight as one from a recognised tool.
That matters because of what happens next. Applying the label burns the marking into the picture and signs the result with our certificate — so an unchecked claim would have become our assertion about your photograph.
From now on the claim is verified before it can make the label available. Where it verifies, nothing changes: you are offered the label exactly as before. Where it cannot be verified, the file is treated as an ordinary upload and the label is not offered.
One distinction we deliberately kept: a file signed by a tool we do not yet recognise still counts. That is a statement about the list of signers we carry on a given day, not about your file, and that list changes as new tools appear. Only a failure that describes the file itself — a broken signature, or content that no longer matches what was signed — withholds the label.
Your own photographs are unaffected. They never carried such a claim and were never offered the label to begin with.