AI created more images in 18 months than cameras did in 149 years

The volume is genuinely hard to wrap your head around. Starling Lab, a research group out of Stanford and USC, points out that cameras needed nearly a century and a half to produce 1.5 billion images. Generative AI hit that same milestone in a year and a half. And it didn’t pause for a breather. Google dropped a number at I/O this spring that resets the scale of the problem: its tools alone have churned out over 100 billion AI images and videos in just a few years. Google is far from the only factory pumping out synthetic media, but that stat explains why the company is now pushing its SynthID watermarking tech with a new sense of urgency.

The pitch is straightforward. SynthID embeds an invisible signal directly into an image’s pixels or an audio file’s waveform. Unlike metadata-based systems like C2PA—which are cryptographically secure but vanish the second you take a screenshot or save a compressed copy—SynthID is designed to be a survivor. Google DeepMind scientist Pushmeet Kohli told Ars that the team worked hard to make the watermark durable enough to stick around through edits, resizing, and the general digital wear-and-tear that memes and viral clips endure. That resilience is now getting a real-world stress test, as OpenAI, Runway, and Nvidia have signed on to use the technology.

I’m of two minds about this. Technical durability is table stakes. The bigger question isn’t whether a watermark can survive a crop, but whether the entire ecosystem will treat a SynthID flag as meaningful. C2PA is a clean, mathematically verifiable standard that breaks instantly in the messy pipeline of social media. SynthID solves the durability problem but introduces opacity. You can’t independently verify it without access to Google’s proprietary detection tools, which turns content authentication into a service you have to trust rather than a fact you can check. That’s a fragile foundation for something as important as separating real from fake.

Still, the sheer scale of the deluge makes some kind of intervention necessary. When a single company’s tools are responsible for 100 billion pieces of synthetic media, the old approaches—forensic analysis, reverse image search—are laughably insufficient. The partnerships with major model builders suggest SynthID might become a de facto standard, not because it’s perfect, but because it’s there and the alternatives are worse. Whether users actually look for these labels, or care about them when they see them, is a cultural problem no amount of engineering can solve.