Identifying AI Watermarks: SynthID, C2PA and the Limits of Verification

Can you tell that an image was created with AI simply by looking at it? Usually not reliably. That is why invisible watermarks, signed provenance data and AI detectors attract so much attention. SynthID and C2PA are two frequently mentioned names. They are often grouped together even though they do different things. This article explains the differences and where their limits begin.

Can a watermark identify an AI-generated image?

AI images now often look so realistic that their origins are difficult to assess reliably with the naked eye. Technical provenance signals are therefore becoming more important. The term “watermark” comes up frequently, but it does not always mean the same thing: some systems add invisible patterns directly to the content, while others sign metadata describing its origin and editing history.

Both approaches help, but neither solves every problem. A verification result can speak only about signals the particular system understands. Failure to detect a signal does not prove a recording is genuine. Likewise, finding a signal does not automatically establish whether the depicted claim is accurate.

What Google SynthID does

Google describes SynthID as a technology for adding invisible markings to AI-generated or AI-edited content. Such signals can occur in supported images, videos, audio files or text. Appropriate verification methods can detect them. Support from the system that generated the content is crucial.

This does not make it possible to verify every image from any source reliably. If another tool adds no compatible signal, or if a file has been heavily altered, a detector may reach its limits. A negative result should therefore be phrased cautiously: “No supported watermark detected” is very different from “Guaranteed not to be AI-generated”.

How C2PA differs

C2PA and the Content Credentials based on it use signed provenance information. This may document which application created or edited a file, for example. Suitable tools can verify the information. Unlike an invisible watermark embedded directly in the content, the primary aim is a traceable chain of origin.

Metadata can disappear during export or conversion. This is one reason researchers are exploring additional robust methods. Even with complete provenance, the editorial question remains: are the crop, description and claimed context accurate? Provenance does not replace checking the substance.

Practical example: Checking a suspicious advertising image

You receive a file that supposedly shows a real event. First check the original source and whether the publication is plausible. Then use suitable tools to look for existing Content Credentials or supported watermarks. Record what the verification report actually confirms and what remains uncertain.

Do not rely on a single detection app. Reverse image search, asking the source directly and checking the circumstances of the recording can also help. If you intend to use the image in advertising, licensing, personality rights and any applicable disclosure obligations also matter. These depend on the relevant jurisdiction; the EU rules linked below are not worldwide rules. When in doubt, creating a reliable illustration of your own may be preferable.

Conclusion

AI watermarks are an important aid, but not a universal truth detector. SynthID and C2PA take different technical approaches, and both have limits. My tip: use provenance signals as one part of a traceable source-checking process, and do not turn an inconclusive result into an absolute claim.

Sources and further reading

The following primary sources and official documentation provide further detail. Always check the current position on legal matters and topics that change over time.

• deepmind.google
• c2pa.org
• iptc.org
• digital-strategy.ec.europa.eu