Google says its tools helped create 100 billion AI images. A watermark can label them, but can it stop the harm?

Key points

• Starling Lab, a research collaboration from Stanford University and the University of Southern California, estimates humanity took 149 years to produce 1.5 billion photographs; generative AI matched that in 18 months.
• Google announced at its I/O developer conference this spring that its tools helped create more than 100 billion AI images and videos in just a couple of years.
• Google is expanding partnerships around SynthID, its technology for tagging AI-generated content, so more platforms can use it.
• SynthID can label AI content at creation, but the label cannot stop fake images from spreading before anyone checks.

Put a number on it: humanity snapped photographs for 149 years before reaching 1.5 billion images. Generative AI, software that produces photos and video from a text prompt, crossed that same threshold in 18 months. Those figures come from Starling Lab, a research collaboration run jointly by Stanford University and the University of Southern California that studies how digital media is authenticated.

Then Google arrived with a bigger number. At its I/O conference this spring, the company said its tools had produced more than 100 billion AI images and videos over just a couple of years. First reported by Ars Technica AI, that figure covers Google's own products alone. Dozens of other companies also run image and video generators.

What is SynthID, and how does it work?

SynthID bakes an invisible watermark, a hidden signal woven into the pixels or audio data, into content the moment Google's systems create it. A separate detector can then read that signal later. The watermark survives a surprising amount of abuse: cropping and screenshot re-saves generally leave it intact.

Google used the I/O announcement to confirm new partnerships that would spread SynthID to more platforms, meaning more AI content created elsewhere could carry the tag. It's the fourth time we've covered AI watermarking since our first report on 22 July, when Meta's rival system failed to flag more than half the images in one real-world test.

So is the problem solved?

Not even close. A watermark's a label, not a lock.

Content carrying a SynthID tag can still spread across feeds and messaging apps long before any detector scans it. Most people have no way to run a check themselves, and platforms aren't required to display a warning even if they could.

There's also everything that never gets tagged. SynthID only marks what Google's own tools produce. Any image made with a rival generator, or with open-source software anyone can run locally, carries no watermark at all.

What should readers watch for?

Labelling technology is genuinely useful. But volume is the real story. When one company's tools can produce 100 billion pieces of synthetic media in two years, the realistic risk isn't one convincing deepfake. It's an ocean of plausible-looking content that overwhelms the human capacity to question it. We've already seen that dynamic play out with fake AI doctors racking up millions of TikTok views with bogus health advice.

A couple of practical habits help. Reverse-image search any photo that seems designed to provoke a strong reaction. If a video shows someone saying something shocking, find a transcript from a named journalist before trusting it.

The watermark's a starting point. Scepticism is still the sharper tool.