Meta built its own AI watermarking system instead of using Google's, and it doesn't work with anyone else's detection tools

Meta Built Its Own AI Watermark. No Other Platform Can Read It.

The technology to detect AI-generated images exists. It works inside each platform that built it. It fails everywhere else, and the gap between those two facts is where fabricated images survive.

Meta introduced Content Seal in July alongside Muse, its new image and video generation model 1. The system embeds an invisible watermark into AI-generated images that persists through cropping, compression, resizing, and screenshots 1. By Meta's own description, Content Seal works similarly to Google's SynthID, an invisible watermarking system that OpenAI has already adopted 1. Meta also operates as a steering committee member of the Coalition for Content Provenance and Authenticity, or C2PA, the non-profit that maintains the Content Credentials specification 1. Existing provenance frameworks from Google and C2PA were already available. Meta built a third.

The Verge ran the test that exposes the cost. A Muse-generated image carrying the Content Seal watermark was submitted to Google's Gemini chatbot and the official C2PA detection portal. Neither tool could confirm the image was AI-generated 1. The only place Content Seal registers today is a dedicated web tool Meta is testing, which carries a daily usage cap that Meta says protects against misuse 1. That tool has not been integrated into the Meta AI chatbot, unlike Google's approach of embedding SynthID detection directly into Gemini 1. Content Seal also covers only images from the Muse model, not Meta's earlier AI systems, and lacks video support 1.

Meta spokesperson Faith Eischen told The Verge the company is "exploring ways to bring detection closer to where people encounter AI-generated content" and is "determined to work with our industry peers" on broader platform support 1. Both commitments are in the future tense. When The Verge asked whether Meta was instructing outside platforms like TikTok and LinkedIn on how to detect Content Seal, Eischen described no active integrations 1.

Meta's own Oversight Board called on the company in March to "meet its public commitments and employ its own tools" to address deceptive generative AI content across its platforms 1. Content Seal is technically that response. But the board asked for tools that reduce the spread of deceptive content. A watermark that only Meta can read, on images from only Meta's newest model, detectable only through Meta's own testing portal, addresses that mandate on paper and undermines it in practice.

The deeper question is why Meta went solo. Eischen said Meta, "like others," built Content Seal "natively towards our own technical specifications and products" 1. But OpenAI did not build its own system. It adopted SynthID 1. C2PA develops open, global technical standards for content provenance 2. The Verge proved that the C2PA portal still cannot detect a Content Seal watermark 1. Building to your own specifications is precisely the behavior a shared standard is designed to prevent.

The consequence lands on anyone who needs to verify an image. A journalist, a voter, a content moderator: all now face a landscape where they must know which AI platform generated an image before they can pick the right detection tool. Knowing the source is the one piece of information they never have. The detection system designed to tell you where an image came from can only function if you already know where it came from. That circularity is not a bug in any single product. It is the predictable result of treating provenance as a competitive asset instead of shared infrastructure.

The barrier is institutional, not technical. OpenAI demonstrated that cross-platform provenance is achievable by adopting Google's SynthID 1. C2PA maintains an open specification for exactly this problem 2, and Meta sits on its steering committee 1. Building a third incompatible watermark was a choice, not a necessity. And until that choice changes, every new standard added to the landscape makes verification harder, not easier.