Google's SynthID for watermarking AI images is hard to break, but watermarking can't change that there will always be ways to AI-generate content without labels
Deciding what's real on the Internet won't be easy in the future. — The scale of AI-generated media can be hard to grasp.
Context & Ripple Effects
Google introduced SynthID as an invisible image watermark for its Imagen customers, then extended the system to text through SynthID Text's developer availability. The effort has since become more interoperable: OpenAI added support for SynthID image watermarks and proposed a verification portal.
This article clarifies the boundary of that progress. A durable watermark can improve attribution for participating tools, but it cannot establish a universal rule for media made by tools that do not apply labels; earlier research also showed watermarks could be removed or falsely added under some conditions.
First-order effects
- Google and participating model providers gain a stronger provenance signal for images they choose to mark, making verification more useful within their own ecosystems.
- Publishers, platforms, and audiences still cannot treat the absence of a SynthID signal as proof that an image is human-made or authentic.
Second-order effects
- Competing model vendors and distribution platforms face pressure to support common provenance signals if they want their generated media to be readily verifiable, while retaining a choice over implementation.
- Moderation and verification workflows must combine watermark checks with other evidence, because label coverage—not only watermark resilience—limits automated decisions.
Third-order effects
- Synthetic-media governance is likely to split between interoperable provenance for cooperative providers and a persistent unlabelled supply of generated content, rather than converge on a single authenticity test.
- If adoption broadens, control over detection and verification interfaces could become as consequential as watermark creation; the corpus does not establish that a universal standard will emerge.
The trend: AI provenance is shifting from a watermarking problem to a coverage and verification-network problem, as robust signals remain voluntary across a fragmented generation market.