Google’s SynthID Works Perfectly. That’s Exactly the Problem.

You’ve seen the deepfake. The politician saying something they never said. The disaster photo that never happened. And you’ve probably thought: someone should fix this.

Well, Google did. They built SynthID, a watermarking system for AI-generated content. And by every technical measure, it works. It’s resilient. It survives cropping, compression, resizing. It’s nearly impossible to strip out.

And it won’t save you. Not even a little.

Here’s the uncomfortable truth nobody in the AI safety crowd wants to say out loud: A perfect watermark on a piece of disinformation is still a piece of disinformation that just got shared 4 million times before anyone bothered to check the label.

The engineers at DeepMind built something elegant. The watermark is imperceptible to humans but detectable by machines. It persists through transformations that would destroy most steganographic tricks. In a lab, this is a triumph. In your aunt’s Facebook feed, it’s invisible.

And that’s the fundamental tension. The very feature that makes SynthID technically brilliant, its imperceptibility, is the same feature that makes it socially useless. A watermark nobody can see is a warning nobody hears.

Think about how misinformation actually spreads. Someone sees a sensational image. They feel outrage. They hit share. They move on. At no point in this chain does anyone pause to run a detection tool, check a provenance ledger, or look for an invisible watermark embedded in pixel noise.

The disinformation pipeline doesn’t fail at detection. It fails at the moment a human feels something and clicks a button before thinking.

One Ars Technica commenter made a point that should haunt every AI policy discussion: swap out “AI” for “Photoshop” and swap “disinformation” for “hoaxes.” These problems have been around forever. We survived Photoshop. We survived video editing. We didn’t survive because we developed better detection tools. We survived because the incentive structures of media eventually adapted, slowly, imperfectly, but they adapted.

Right now, the incentive structures of social platforms reward engagement above all else. A fake image of a burning city generates more ad revenue than a true image of a boring city. The algorithm doesn’t care about watermarks. It cares about clicks.

SynthID is a band-aid on a systemic wound. It’s a brilliant band-aid, engineered with extraordinary care, but it’s covering a hemorrhage.

You cannot solve a social problem with a technical solution when the social problem is that nobody is incentivized to care.

The real leverage point isn’t detection. It’s accountability. It’s forcing platforms to bear the cost of the harm their engagement-optimized systems cause. It’s making it expensive, legally and reputationally, to amplify manipulated content at scale.

But that’s hard. That requires regulation. That requires political will. That requires admitting that the free-market utopia of “the internet will self-correct” was always a fairy tale told by people who profited from the chaos.

So instead, we get watermarks. Invisible ones. That technically work. That nobody will ever check.

And the next deepfake will still go viral. And the next politician will still deny saying what they said. And the next disaster photo will still get shared 4 million times.

Because the problem was never detection. The problem was always that we built a machine that profits from lies and then asked it nicely to stop.