Meta's invisible watermark: what Meta checks on Facebook, Instagram and Threads
Meta says photorealistic images made with Meta AI carry invisible watermarks and IPTC metadata, and that it labels other companies' images when it finds industry-standard signals in them. Meta also says people can strip out invisible markers. Our checker does not read Meta's watermark, so it can neither confirm nor rule one out.
Does Meta add an invisible watermark to AI images?
Meta says yes, for its own tool. In its February 6, 2024 post on labeling AI-generated images (the page shows an update dated April 1, 2025), Meta says that photorealistic images made with its Meta AI feature get visible markers, invisible watermarks and metadata embedded in the file. It names the invisible markers as IPTC metadata and invisible watermarks, and says using both improves the robustness of the invisible markers and helps other platforms identify them.
That post does not name the watermark or say how it works. It also describes a watermarking technology Meta calls Stable Signature, which Meta says it is developing. It does not say that method is in use.
What does Meta say it checks on Facebook, Instagram and Threads?
The same February 2024 post says Meta would label images that people post to Facebook, Instagram and Threads when it can detect industry-standard indicators. It says it was building tools to read the "AI generated" information in the C2PA and IPTC standards, so that it can label images from Google, OpenAI, Microsoft, Adobe, Midjourney and Shutterstock as they implement their plans for adding metadata. It adds that it could not yet detect such signals in audio and video from other companies.
Meta's April 5, 2024 post gives two triggers for a label: it detects "industry standard AI image indicators", or the person posting discloses that the content was made with AI. It also records that "Made with AI" became "AI info" on July 1, 2024. Our guide to Meta's AI info label covers it.
What is Meta's Content Seal?
Meta's July 7, 2026 post "Introducing Muse Image and Muse Video" says Muse Image includes Content Seal, which Meta calls its invisible watermarking system. Meta says images created by Muse Image in the Meta AI app and on meta.ai carry a hidden provenance signal that stays intact "even when cropped, compressed, resized, or screenshotted". It says it plans to extend Content Seal to video, and that it is previewing a detection tool for the watermark.
Some further details are reported, not stated on Meta's pages. Engadget reported on July 7, 2026 that the tool at meta.ai/identification covers only images created or edited with Muse Image, and that it did not identify images from earlier Meta AI model versions in the author's testing. The same report says Content Seal is not compatible with SynthID or C2PA Content Credentials, and that Meta's new models do not add visible watermarks.
We also opened the identification page itself on October 30, 2026. Its heading reads "Identify files created or edited with AI". It invites you to upload an image, video or audio file, and it shows Log in and Sign up buttons. It did not mention Content Seal, Muse Image or any limit.
What does Meta say, side by side?
| Topic | What is said | Source |
|---|---|---|
| Photorealistic Meta AI images | Visible markers, invisible watermarks and embedded metadata (IPTC). | Meta, February 6, 2024 |
| Other companies' images | Meta said it would label them when it detects industry-standard indicators such as C2PA and IPTC information. | Meta, February 6, 2024 |
| Label triggers | Detected industry-standard indicators, or the poster's own disclosure. | Meta, April 5, 2024 |
| Content Seal | An invisible watermark in Muse Image output, with a detection tool in preview. | Meta, July 7, 2026 |
| Content Seal tool limits | Reported: covers only Muse Image; older Meta AI model versions were not identified in the author's testing. | Engadget, July 7, 2026 (reported) |
| Limits | Not all AI-generated content can be identified; people can strip out invisible markers. | Meta, February 6, 2024 |
What limits does Meta admit?
Meta's February 2024 post says it was not yet possible to identify all AI-generated content, and that there are ways for people to strip out invisible markers. It says Meta was working on classifiers for content without markers, and on ways to make invisible watermarks harder to remove. It calls this likely to become an increasingly adversarial space.
So a label, or no label, on a Facebook, Instagram or Threads post says little on its own. Meta's April 2024 post ties the label to a signal Meta detected or to something the poster said. A missing label may only mean that no signal survived.
What did we find in files filed under Meta AI?
We hold two real files filed in the Wikimedia Commons category "Images generated by Meta AI". Both are WebP files of 1,280 by 1,280 pixels. Third parties uploaded them, and we cannot know how each was handled before upload, so they show what our checker reads, not what Meta adds.
Our measured result: 2 files, measured October 27, 2026; the Content Credentials result was measured again on October 30, 2026 and did not change. Neither carries a Content Credentials record or an IPTC source-type declaration. We found no XMP, EXIF, IPTC or text metadata in either. The checks for Content Credentials, metadata and the Stable Diffusion watermark all ran, and both files read "No supported provenance signals found". The counts are on our accuracy page.
Our checker does not read Meta's watermark. These results therefore cannot say whether either file carries one, or whether Meta's tools would label either file.
What this means when you check an image
Open the checker, choose the file, and look for the headline "No supported provenance signals found". It means we found no signed record, no source label in the metadata and no Stable Diffusion watermark pattern. The line under it says this does not mean the image is real or human-made. If a file still carried an unsigned source-type label in its metadata that names a generative source, the headline would read "Metadata declares AI use" instead, and the detail would say it is a label, not proof.
Either way, the result lists Meta's watermark under what is not checked here. A copy saved from a post may have lost signals that Meta read on the original, and Meta may use signals we do not.
If you tick our box that reads I can see an "Imagined with AI" label, or the file's metadata names Meta AI, the checker lists a link to Meta's identification page, which Meta describes as a preview. It opens another company's site, you may need to sign in, and your image is uploaded to them, not to us.
Guides on this topic
- What is SynthID? is the main page on invisible watermarks.
- Meta's "AI info" label covers what triggers the label and what it does not prove.
- Invisible AI watermarks compared sets Meta's mark beside SynthID, TrustMark and the Stable Diffusion mark.
- Can AI watermarks be removed? covers what survives cropping, screenshots and re-saving.
- Meta's AI info labels on ads covers the label on advertising.
Frequently asked questions
- Does Meta add a visible watermark to AI images?
- For photorealistic images made with Meta AI, Meta's February 2024 post says it puts visible markers on the images. Engadget reports that Meta's new models do not add visible watermarks, unlike some earlier Meta AI versions. We cannot tell which applies to a given picture.
- Can our checker read Meta's invisible watermark?
- No. The only invisible watermark it decodes is the one Stable Diffusion's reference software adds. The result lists Meta's watermark as not checked.
- Does Meta label every AI image on Facebook, Instagram and Threads?
- Meta does not say so. Its pages say it labels what it detects through industry-standard indicators or what people disclose, and that it cannot identify all AI-generated content.