Meta Built a Tool to Spot Its Own AI-Generated Images—With a Catch

Meta has released a detector tool that identifies images created or edited with Muse Image, its newest image-generation model. You can try it now at meta.ai/identification. Engadget

The detector works by looking for an invisible watermark that Meta embeds into images using a system called Content Seal. Think of this watermark as a hidden fingerprint: it's baked into the image data itself, not visible to the eye. According to Meta's announcement at ai.meta.com, this watermark is designed to withstand common image manipulations—cropping, compression, resizing, even screenshots of screenshots—without disappearing. That's the critical engineering claim. Most watermarks fall apart when an image gets recompressed or resized, so if Content Seal truly holds up under those conditions, it addresses a real technical problem.

Unlike earlier Meta AI image tools, Muse Image doesn't add a visible logo or watermark to generated images. That's a deliberate trade-off: there's no visual badge a person can glance at to flag synthetic content, which means everything depends on the invisible watermark actually working as advertised.

For now, the detection tool only works on images made with Muse Image. Meta says it plans to extend watermarking to AI-generated and edited video, and the company is developing a video generation model called Muse Video, described as "coming soon." Until video watermarking launches, you can't use this tool to verify video provenance.

The Technical Limits

Meta has previously released open-source watermarking research, but the version tied to Muse Image is proprietary and closed off. Outside researchers can't independently verify how robust the watermark actually is against someone trying to remove it—they can only test results through Meta's own detection system. That's a limitation worth understanding: independent security audits catch problems that vendor testing sometimes misses.

In testing reported by Engadget, the web tool failed to identify images created with earlier versions of Meta's AI image models. So the detector doesn't retroactively cover the substantial volume of AI-generated images Meta's older tools have already produced. Engadget

The detector also has a daily usage cap—you can't run unlimited checks—and it only exists as a standalone web page right now, not built into the Meta AI app itself.

The Ecosystem Problem

Here's where the story gets wider. Content Seal doesn't work with SynthID, Google's watermarking system, or with C2PA Content Credentials, which is a shared industry standard backed by Adobe, Microsoft, OpenAI, and others. A file watermarked with Content Seal won't show as watermarked in a SynthID checker, and vice versa. Platforms and researchers wanting to verify images at scale now need to maintain multiple, separate detection systems instead of using one common tool.

C2PA was built specifically to be a universal format—any company using it should be able to read any C2PA watermark, regardless of who created the content. SynthID has been adopted across Google's products to build that kind of cross-platform trust. Meta's choice to build a proprietary, non-interoperable watermark cuts against that convergence. Whether that's a deliberate competitive advantage or simply a product built on Meta's own schedule isn't something the available facts answer.

The deeper issue here is that watermarks only matter if they work everywhere content actually travels. Meta's own detection tool works fine for images checked directly against it. But the real-world problem watermarking is meant to solve—misinformation and AI-generated media spreading across platforms and being checked by regular people—happens on the open web. A closed detection tool that only Meta controls solves that problem only if people route their verification back through meta.ai, which most won't do at scale.

What Comes Next

Meta has stated clearly that video watermarking will arrive once Muse Video ships, and the detection tool will likely move from a standalone web page into the Meta AI app. The more consequential question is whether Content Seal eventually connects with the broader C2PA ecosystem or remains a separate, Meta-specific track. That choice will tell us a lot about what AI provenance standards look like in the next few years.