AI Image Watermarks Actually Work Now
For years watermarking was a polite suggestion. SynthID Bio and friends changed the math.
Watermarking AI images used to be a joke you could strip in two clicks. The new generation is different. Google's SynthID Bio, a watermarking system for AI-designed proteins, reported 100% detection at a 0.1% false-positive rate in a Nature paper this month, which is a different league from the old visible overlays.
The shift matters because the old debate was 'can we tell' and the answer was usually no. Now the answer for some domains is yes, reliably, even after edits. That changes what platforms can enforce and what users can trust when they see a generated image.
It's not universal yet. Text and video watermarking lag, and adversarial edits still beat some schemes. But the direction is clear: provenance is becoming a property of the file, not a label someone forgets to add.
For anyone publishing AI visuals, the takeaway is to use tools that embed provenance by default. The era of 'looks real enough' is ending; the era of 'provably generated' is starting, and that's healthier for everyone except the bad actors.