Meta has built an AI detection tool to identify images and video generated with Muse Image, its newest image model, and is previewing the tool as a web-based checker at meta.ai/identification. Engadget

The detector works by scanning uploaded media for invisible watermarks embedded through a system Meta calls Content Seal. According to Meta's own blog post announcing the tool, ai.meta.com, the watermark is designed to survive common forms of manipulation — cropping, compression, resizing, even a screenshot of a screenshot — without losing detectability. That durability claim is the operative technical bet here: most watermarking schemes degrade badly once an image passes through a few rounds of social-platform recompression, so Meta is asserting Content Seal holds up specifically in the conditions where provenance signals usually die.

Muse Image ships with no visible watermark. That's a reversal from earlier Meta AI image tools, which stamped a small logo in the bottom right corner of generated output. The tradeoff is now entirely dependent on the invisible layer working as advertised, since there's no visual cue left for a human glancing at an image to flag it as synthetic.

Detection currently only covers images created or edited with Muse Image. Meta says it plans to extend Content Seal watermarking to AI-generated and edited video, and the company is separately developing a video generation model, Muse Video, described as "coming soon." Until that ships, video provenance checking through this tool doesn't exist.

The proprietary version of Content Seal embedded in Muse Image is not the same as the versions Meta has previously open-sourced. Meta has released open-source watermarking work before, but the implementation tied to its newest models is closed, which means outside researchers can't independently audit how the watermark is generated or how robust it actually is against adversarial removal — they can only test the output through Meta's own detection endpoint.

That endpoint has real gaps. In Engadget's testing, the web tool failed to identify images that had been created or edited with earlier versions of Meta's AI image models — meaning Content Seal detection doesn't retroactively cover the substantial body of AI imagery Meta's tools have already produced. Engadget The detector is also rate-limited, with a daily cap on how many identification checks a user can run, and it is not yet integrated into the Meta AI app itself — it exists only as a standalone web page for now.

Perhaps the more consequential detail for the broader ecosystem: Content Seal is not interoperable with SynthID, Google DeepMind's watermarking scheme, or with C2PA Content Credentials, the cross-industry provenance standard backed by Adobe, Microsoft, OpenAI, and others. A file watermarked with Content Seal won't register on a SynthID checker, and vice versa. Platforms and researchers trying to verify provenance at scale now have to juggle multiple, non-communicating detection systems rather than one shared standard.

The fragmentation here is worth flagging on its own terms. C2PA was built explicitly to be a common format that any signatory could read regardless of which company generated the media, and SynthID has been adopted across a growing set of Google products specifically to build that kind of cross-platform trust. Meta choosing a proprietary, non-interoperable watermark — while still participating in some industry provenance conversations — pushes against that convergence rather than toward it. Whether that's a deliberate competitive moat or simply a product built on Meta's own timeline is not something the available facts settle.

In this author's view, the gap between "detectable when checked directly against Meta's own tool" and "detectable across the open web where content actually spreads" is the real test Content Seal hasn't yet passed. A watermark that survives compression and cropping is a meaningful engineering achievement if the claim holds up under independent scrutiny. But provenance tools only matter at the scale of the problem they're meant to solve — misinformation and synthetic media moving across platforms Meta doesn't control, checked by people using tools Meta doesn't build. A closed, single-vendor detector that can't talk to SynthID or C2PA solves that problem only for the slice of the internet willing to route everything back through meta.ai.

The near-term trajectory is clear enough from Meta's own statements: video watermarking is coming once Muse Video ships, and the detection tool will presumably move from a standalone web preview into the Meta AI app proper. Whether Content Seal eventually finds a bridge to the broader C2PA ecosystem, or remains a parallel, Meta-specific track, will likely say more about the future shape of AI provenance standards than the watermarking technology itself does.