Which image records survive a platform upload?

Results: what X kept

What else we measured

• PNG uploads stayed PNG. All 3 served PNGs were byte-identical to the same image with no records (SHA-256 acdc3a0a5e01…): X dropped the record chunks and kept the image data unchanged.
• JPEG uploads stayed JPEG at the same dimensions, with identical decoded pixels. X rewrote them as progressive JPEGs and removed the EXIF and XMP segments.
• The default URL, ?name=large and ?name=orig returned the same bytes for every file.
• The C2PA reader reported NO_C2PA_MANIFEST for the downloaded copy of the signed file. A missing manifest after upload says nothing about whether the original was AI-generated.

Platforms tested and still to test

Each platform will be tested with the same five files and added here with its date. Platform-specific background is on the AI image checks by platform pages.

Method

1. We made five 640×400 test images from one code-drawn picture and attached one known record to each: a PNG "parameters" text record in the AUTOMATIC1111 format, a PNG with ComfyUI editor and API graphs, a JPEG with an XMP packet declaring IPTC DigitalSourceType trainedAlgorithmicMedia, a JPEG with invented EXIF camera and GPS fields, and a PNG with a C2PA manifest signed by a disposable test certificate. The records are test data, not evidence of how the pixels were made.
2. Before uploading, we read every file with the same isGenAI code the site's tools use: the C2PA checker, the IPTC/XMP AI source reader, the EXIF viewer, the saved prompt extractor and the ComfyUI workflow reader. We recorded each file's SHA-256.
3. On 19 October 2026 we uploaded the files with the X API v2 media upload (category tweet_image) and attached them to 2 posts from @isgenai that said they were test uploads for this study. We fetched each post's media URL from the X API, downloaded the served file from pbs.twimg.com in three variants (default, ?name=large, ?name=orig) and then deleted the posts.
4. We read each downloaded copy with the same readers and compared the records, format, dimensions, byte size, SHA-256 and decoded RGBA pixels.

Download the files

Everything is released under CC0-1.0: use it anywhere, no permission needed.

• results.json: every downloaded copy with hashes, dimensions, sizes and record-by-record survival
• fixtures.json: the test files, their sources, hashes and "before" readings
• x-run.json: X post and media ids, served URLs, HTTP status and deletion times
• readme.txt: description and reproduction steps
• a1111-parameters.png: PNG tEXt "parameters" record (prompt, negative prompt, steps, sampler, seed, model). SHA-256 349c5489f092e0045e79c207acdfe16e217ea7dadf59f2512db6aeccf725aff5
• comfyui-workflow.png: PNG "workflow" (editor graph) and "prompt" (API graph) records. SHA-256 427aa019b768aaaeb2fb59724a4e49772e041ee6091f47cffc7ec6a35867dc08
• iptc-ai-source.jpg: XMP Iptc4xmpExt:DigitalSourceType = trainedAlgorithmicMedia. SHA-256 ea6647c57631c171a72ec457f04272dedee91919238629275eff4987f5518504
• exif-camera-gps.jpg: EXIF Make/Model, DateTimeOriginal, exposure settings and GPS latitude/longitude (invented values). SHA-256 8cad735afdc552596c4bdb170b31df29bb3f6735f82c28fcf9060f91c813209e
• c2pa-signed.png: C2PA manifest signed with a disposable test certificate; c2pa.created action with IPTC algorithmicMedia. SHA-256 ce4b8329b838e06147f2c8b002ee366d23bcdecb9bba4b322481920944c255c5

Downloaded X copies are in /research/platform-metadata/x/, for example c2pa-signed.orig.png. The study script is scripts/platform-metadata-study.ts in the isGenAI source.

Cite this data

The data is CC0, so credit is not required, but a citation or link helps readers find the files, hashes and test date behind any figure you quote.

isGenAI. (2026, October 19). Platform metadata survival study: What X keeps [Data set]. https://isgenai.com/research/platform-metadata-survival

@misc{isgenai_platform_metadata_2026,
  author       = {{isGenAI}},
  title        = {Platform Metadata Survival Study: What X Keeps},
  year         = {2026},
  month        = oct,
  howpublished = {Data set, CC0-1.0},
  url          = {https://isgenai.com/research/platform-metadata-survival},
  note         = {X tested 2026-10-19; data: https://isgenai.com/research/platform-metadata/results.json}
}

<a href="https://isgenai.com/research/platform-metadata-survival"><img src="https://isgenai.com/research/platform-metadata/badge.svg" width="220" height="40" alt="Data: isGenAI platform metadata survival study (CC0)"></a>
<p>Data: <a href="https://isgenai.com/research/platform-metadata-survival">isGenAI platform metadata survival study</a>, CC0.</p>

Limitations

• One upload path per platform. We used X's official API; the web uploader, the mobile apps and direct messages may process files differently.
• Small synthetic images. Large camera photos, photos with transparency, GIFs and videos may be resized or re-encoded differently.
• One date. Platforms change their processing without notice, so each result carries its test date and file hashes.
• Records, not origin. A record that disappears says nothing about whether an image is AI-generated, and a record that survives can still be edited by anyone. Read how to check an image's history for what each record proves.

Compare with our local export experiment and the Image Metadata Lab. Report a problem with a file or result.