A feature image we made with Google Gemini for one of our own posts. We posted it to LinkedIn, Mastodon and X, downloaded the copy each one served back, and checked all four.
What changed
- The label. The original carried Content Credentials signed by Google, declaring it AI-generated. None of the three copies had them.
- The picture, sometimes. LinkedIn made it smaller and Mastodon recompressed it. X sent back the same pixels and only removed the label.
- The estimate never caught it. 3% on the original, 1 to 3% on the copies. Without the label, the report called every copy "likely authentic".
What the report said: Authenticity F for the original, certified AI-generated. D for the LinkedIn and X copies, B for Mastodon's.
A missing label proves little: all three of these platforms removed it.
Updated 15 October 2026
We checked all four files again. The findings above are what the report said on 9 October. The grades have not moved: F for the original, D for the LinkedIn and X copies, B for Mastodon's.
- The original reads F "AI-Generated (Certified)", decided by its signed label.
- Every copy now reads C "No AI-Generation Evidence Found", not "likely authentic". The estimate (1 to 2%) is a weak line in the evidence table.
- The pixel analysis did not catch them. It found no signs in any of the three copies, including X's, which has the original's pixels. It found none in the original either; the label decided that one.
A C is not a finding that an image is real. What each card and reading means: How to read your forensic report.
Go deeper
- How to check an image yourself: Is this photo real or AI-generated?
- What a label can and can't prove: What C2PA Content Credentials Prove, by Kenneth Springer
- Try it on your own image
Tagged
ai-images
