Why Content Credentials disappear: screenshots, messaging apps, social uploads and conversion

A Content Credentials record is often stored inside the file, so anything that writes a new file can leave it behind. OpenAI says everyday actions such as a screenshot or a file conversion can lose it. In our test, none of 9 signed files kept its record when re-rendered. A missing record does not show where an image came from.

Why do Content Credentials disappear?

A record is not part of the picture. It is information attached to the file. OpenAI's help article on provenance signals says platforms, editing tools and file conversions may remove metadata, and that Content Credentials can be lost through everyday actions such as taking a screenshot or converting a file.

The standard says the same thing in general terms. The C2PA explainer asks whether provenance metadata can be removed and answers "Yes it can." The Content Authenticity Initiative's post on durable Content Credentials adds that metadata of any kind can be removed deliberately or accidentally.

Nobody has to mean to remove it. A record that sits in the file can be left behind when the file is written again, and a later section shows what that looks like in our own files.

What do the vendors say about uploads and sharing?

The TikTok post and Midjourney's page are statements by those companies. We have not run an upload through any platform or messaging app, so we do not say which one removes what. For the platforms that read or show a record, see which platforms read or show Content Credentials.

What happens to a record when a file is re-rendered?

We tested the simplest case. We took the 9 real files in our research set that carry a Content Credentials record: 3 Adobe Firefly, 2 Microsoft Image Creator, 3 ChatGPT (two filed on Wikimedia Commons and one sample file) and the C2PA test image. On October 30, 2026 we decoded each in headless Chrome, drew it on a canvas of the same size and wrote it out again, once as a PNG and once as a JPEG at quality 0.92. This is like what a screenshot, a browser export or a re-encode does: the pixels are kept and the file's own data is not.

Result: 0 of 9 files kept a record after the PNG re-render, and 0 of 9 after the JPEG re-render. All 18 outputs read "No supported provenance signals found", so no metadata was found either.

The limits matter. This was one browser and one JPEG quality, with no resize. It was a canvas, not a platform, and we contacted no platform. A canvas holds pixels only, so it shows that a record does not travel through that kind of re-draw. It does not show what any app or site does. We keep the per-file results in our measurement files. Our accuracy page lists what the checker cannot see, including a record that has been stripped.

What did we find in the ChatGPT files we hold?

We hold three files filed in the Wikimedia Commons category "Images generated by ChatGPT". Third parties uploaded them, and we cannot know how each was handled before upload. We measured them on October 30, 2026.

• Two PNG files carry a record. One reads "Trusted Content Credentials declare AI use". The other, an older file whose signing certificate has expired, reads "Content Credentials found; signer trust not established".
• The one JPEG carries no record. It holds only EXIF dates and a colour profile, and reads "Metadata found; source not established".

That is what we hold, and we cannot say why the JPEG has no record. It could have been exported another way, converted, or handled by a site. The files do not tell us, and they say nothing about what any product or platform does.

What if a record points to an online copy?

A file can carry a link to a record stored elsewhere instead of holding the record itself. We made test files like this in JPEG, PNG, WebP, AVIF and GIF. All five read "Content Credentials are not available locally". We never fetch the linked record. In a test with the JPEG, a canary server on the address the file pointed to received 0 requests, so the check did not ask for the linked record. The line under that headline points to Adobe Inspect for a lookup.

The soft binding idea is the other route back. The Content Authenticity Initiative's 2024 post describes decoding a watermark, looking up the record on a cloud host or ledger, and confirming the match with a fingerprint. We do not do that lookup, so our checker cannot recover a record that is gone from the file. Our guide to durable Content Credentials explains the idea.

What this means when you check an image

Open the checker, choose the file, and look for the headline "No supported provenance signals found". It means we found no signed record, no source label in the metadata and no Stable Diffusion watermark pattern. The line under it says: "This does not mean the image is real or human-made. Most images lose these signals when they are shared, screenshotted or re-saved."

It does not tell you where the image came from or that a record ever existed. If the headline is "Content Credentials are not available locally", the file points to a record stored online, and we have not read it. Neither result is a verdict on the picture. The checker also always shows: "These checks cannot determine whether an image is AI-generated."

If you have the original export, check that file, not a screenshot or a copy from a chat or a feed. OpenAI gives the same advice for its own checker. Our Privacy and method page explains what we do with your file.

Guides on this topic

• What are Content Credentials? is the main page on this topic.
• Durable Content Credentials explains how a watermark and a fingerprint can point back to a record.
• Does a screenshot keep metadata or Content Credentials? looks at screenshots on their own.
• Does sharing an image remove its metadata? covers what vendors say about stripping and what we measured when we re-saved real files.
• Which platforms read or show Content Credentials? covers what LinkedIn, TikTok, YouTube, Meta and Google say about reading or showing a record.

Frequently asked questions

Do Content Credentials survive upload?

It depends on the route. OpenAI and the C2PA explainer say a record can be removed, and TikTok says its AI-generated content labels may be removed when content is reuploaded or edited elsewhere. We have not tested any platform upload, so we cannot say which ones keep a record.

Does a screenshot keep Content Credentials?

OpenAI says a screenshot can lose them. In our canvas test, 0 of 9 records survived a re-render, and we describe the limits above.

If no record is found, was the image edited?

We cannot tell. A missing record can come from a screenshot, a conversion, an upload, or a file that never had one.

Can a lost record be found again?

The Content Authenticity Initiative describes a watermark and fingerprint lookup that can do this. Our checker does not run that lookup.