C2PA, EXIF and XMP: what image metadata reveals

01The three formats

EXIF is the oldest and the most personally revealing. Cameras and phones write the make and model, the exact date and time, exposure settings, and — unless location services were disabled — the GPS coordinates where the shutter fired. A photograph shared straight off a phone can disclose a home address to anyone who opens the file properties.

XMP is Adobe's extensible metadata standard, carrying editing history, captions, keywords and rights information. It typically records which software touched the file and when.

C2PA is the newest. Developed by the Coalition for Content Provenance and Authenticity, it attaches a cryptographically signed manifest describing how a file was produced, including whether a generative AI system created or modified it. Because the manifest is signed, tampering is detectable — the signature stops validating.

02Why re-encoding destroys all three

All three formats live in container segments alongside the compressed image data rather than inside the pixels. A JPEG stores them in application markers — APP1 for EXIF and XMP, APP11 for C2PA — which sit before the pixel data in the file.

Drawing an image onto an HTML canvas and exporting it produces a file built solely from pixel values. The decoder reads the source, hands over a bitmap, and the encoder writes a fresh file. No metadata segment survives, because nothing carries it across.

For C2PA specifically the signature makes this irreversible. The manifest is signed over the file's contents, so a re-encoded file cannot present a valid manifest even if one were reattached — the bytes it attests to no longer exist.

03What re-encoding does not remove

Anything carried in the pixels themselves. Some provenance systems pair the signed manifest with a watermark embedded directly in the image data, designed to survive re-encoding, cropping and screenshotting. Re-encoding removes the container and leaves that mark untouched.

Provenance systems may also identify content by perceptual hash — a fingerprint of what the image looks like rather than its exact bytes. Because a re-encoded image looks the same, it still matches, and the original record can be retrieved from a database without any metadata being present in the file at all.

The honest summary is that stripping metadata removes what the file says about itself. It does not remove what the picture itself reveals.

04A note on quality

Re-encoding a JPEG is lossy. The image is decompressed and recompressed, and a small amount of detail is lost each time. At high quality settings this is not visible in normal viewing, but it is real, and it accumulates if a file is processed repeatedly.

Converting between formats compounds this differently: a JPEG re-encoded as PNG loses nothing further but typically becomes several times larger, because PNG is lossless and poorly suited to photographic content. Matching the output format to the input avoids the surprise.

Questions