AI image detector by platform

Discord AI image detector: verify the attachment, not the preview

Discord publishes no policy on image metadata, its previews are re-encoded, and users have reported PNG text metadata being removed from attachments, so a Discord image should be downloaded as the original attachment and verified against the sender’s hash before isGenAI’s report is trusted as the original’s.

Controlled example PNG file that carries this record: No record in this copy

without-records.pngPNG file · controlled example

  1. Signature + IHDR
  2. iTXt XML:com.adobe.xmp
  3. tEXt parameters
  4. tEXt prompt / workflow
  5. tEXt invokeai_metadata
  6. eXIf
  7. caBX C2PA manifest
  8. IDAT pixels
  9. IEND

No record in this copy

Screenshots and re-exports often drop records.

A clean copy is not proof of a human author, so the report says what is missing and asks for the original.

Check an image →
Records like these are often removed when a picture passes through Discord, so the report says what is missing instead of guessing. The picture is a code-drawn controlled example with no records; download it and check it yourself.

What Discord strips and adds

Discord shows re-encoded previews in chat and serves attachments from its CDN. Posts in Discord’s own support forum report text metadata disappearing from PNG attachments and trailing data being removed, which broke communities that share AI character cards and Stable Diffusion workflows inside PNGs. Discord has not published a statement on this as of October 2026, so behaviour may differ by client, size and date. Discord applies no AI label.

Midjourney images delivered through Discord are previews of files generated elsewhere; the Midjourney website download is the original.

What isGenAI shows for a Discord image

The AI image detector reads the file you upload: a C2PA manifest if one survived, the IPTC DigitalSourceType in XMP, AI statements in metadata, saved prompts in the AUTOMATIC1111, ComfyUI and InvokeAI formats, and matching copies by hash and perceptual fingerprint. A platform copy with none of these returns “no evidence found”, and the report says plainly that absence proves nothing about the author.

Compare the attachment’s SHA-256 with the sender’s copy using the image SHA-256 checksum tool. A match proves you have the original bytes; a mismatch means Discord or a client changed the file.

What to do instead

Download the attachment through the original-file link rather than saving the preview, and ask the sender to share a hash or to upload the file as a zip, which Discord does not alter. The image metadata compare tool shows which records a changed copy lost.

For Midjourney and ComfyUI images, the generator pages explain what the originals contain.

How to check an image from this platform

  1. Note the platform’s own label or credential icon on the post before you download anything; it does not travel with the file.
  2. Ask the poster for the original file, or trace the first upload. A platform download is a re-encoded copy on most services.
  3. Upload what you have to the AI image detector and read the report. A stripped copy returns “no evidence found” honestly.
  4. Use the matching-copy check: a hash or perceptual fingerprint match can tie the stripped copy to an original that still carried records.

Check the downloaded image

Frequently asked questions

Does Discord remove image metadata?

Discord publishes no policy. Users report PNG text metadata being removed from attachments and previews are re-encoded. Verify the download’s hash against the sender’s file.

Can isGenAI read a ComfyUI workflow from a Discord PNG?

Only if the attachment kept its text chunks. A zipped file or a hash-verified download is the safe route.

Are Midjourney images on Discord originals?

No. They are delivered previews; download the original from the Midjourney website or ask the creator for it.

Does Discord label AI images?

No platform label as of October 2026.

Sources

Where a platform documents its behaviour, the vendor page is linked. Where it does not, the page says the behaviour is reported, not documented.