How Meta’s “AI Info” Label Works on Instagram, Facebook, and Threads
Meta labels posts from “industry-shared signals” and self-disclosure. That explains both what triggers “AI info” and why lightly edited real photos kept getting it.
11 practical guides about provenance.
Meta labels posts from “industry-shared signals” and self-disclosure. That explains both what triggers “AI info” and why lightly edited real photos kept getting it.
TikTok labels AI content from three sources: your toggle, C2PA Content Credentials, and a watermark only TikTok can read. Here is how each one works.
A post gets the label one of two ways: you add it, or X adds it at upload and locks it. Here is what is documented about each path, and what is still a black box.
One XMP field decides whether a file says “created with generative AI,” “edited with generative AI,” or “captured by a camera.” Here is the vocabulary and who reads it.
LinkedIn’s CR icon is not an AI label. It is a viewer for whatever the file’s signed Content Credentials declare, AI or camera, and it only appears when they survive.
YouTube asks you to disclose realistic AI content, then labels some videos itself. C2PA metadata and YouTube’s own AI tools make that label permanent.
Images from ChatGPT and DALL·E usually carry no visible mark. What they carry is signed C2PA provenance metadata, and newer ones may also carry an invisible pixel signal.
Content Credentials are container metadata, not pixels, which makes them comparatively easy to remove and easy to verify once they are gone.
Provenance can live in visible pixels, invisible signals, file fields, signed credentials, and external records. No single export tells the whole story.
The sparkle and the “Made with Google AI” label are visible pixels. The provenance that ships with them is not. Here is the difference.
An AI label can appear in the X composer when a file carries provenance information. Here is what that means and what it does not prove.