If you screenshot an AI image, what provenance survives?
A screenshot is a new file. That one sentence explains why some provenance disappears instantly while other signals may still be detectable.
People often talk about “the watermark” as if AI images carry one universal stamp. They do not. A file can contain several kinds of evidence at once, and those layers react differently when you screenshot, crop, resize or re-encode the image.
A screenshot breaks the original file container
C2PA Content Credentials live with the media file as signed provenance data. Take a screenshot and you are no longer holding that original file. You created a new image from the pixels displayed on screen, so the original container-level credential is generally not copied into the screenshot automatically.
The same logic applies to ordinary EXIF or XMP metadata. A screenshot application may add its own fields, but it does not normally recreate the source file’s full metadata history.
Pixel-level watermarks are a different question
Invisible watermarking systems such as SynthID are designed to embed a signal into the generated content itself rather than only into the file wrapper. That makes the question more interesting: some transforms may preserve enough of the signal for a supported detector to recognize it.
OpenAI currently describes its provenance approach as layered: supported images can carry both Content Credentials and SynthID, and its verifier can check supported files for those signals. OpenAI also avoids concluding “not AI” when neither signal is found, because provenance can be lost. See OpenAI’s provenance overview and Verify.
So what does a screenshot prove?
Very little by itself. If the screenshot has no C2PA manifest, that tells you the screenshot does not carry the original credential. It does not tell you whether the source image had one before the screenshot was taken.
If a provider-specific watermark is still detected, that is much more useful: it means the screenshot still contains a supported signal associated with that provider. If no watermark is detected, the safest reading is still narrow: no supported signal was found in this copy.
The order of operations matters
If provenance matters, verify the original before you edit it. Save the source file, hash it if needed, inspect the credential, then make a sharing copy. Once the only copy you have is a screenshot downloaded from a chat or social feed, some of the strongest evidence may already be gone.
This is also why “AI detector” should not be one giant green or red badge. A useful checker should tell you which layer it checked: signed provenance, ordinary metadata, provider watermark or a probabilistic visual classifier.
Check the file before transforming it
If you still have the original, inspect that first. A derivative can be useful, but it may have already thrown away the provenance you wanted to verify.
Read the C2PA guide →