SynthID Bio Watermarked Proteins. Provenance Is Now a Science Problem.
Google's protein watermark hits 100% detection at a 0.1% false-positive rate. Labeling AI output just became measurable, not aspirational.
SynthID Bio, a watermark for AI-designed proteins, reached 100% detection at a 0.1% false-positive rate in a Nature paper. That's a step past 'we should label AI output' into 'we can, reliably, even on molecules.' Provenance stopped being a slogan.
The EU's transparency rule already demands labels on AI content, but detection at this accuracy is what makes the rule enforceable. A watermark you can't find is a promise; one you can verify is a control. Bio is the hard case, and it worked, which suggests the technique generalizes.
For builders, the direction is clear: assume generated output will be checkable, not just disclosable. Designing for provenance now, embedding a signal at creation, is cheaper than bolting on a label after the fact when a regulator asks.
The interesting edge is intent. A watermark proves origin, not harm, so the science settles 'who made this,' and the law still has to settle 'was that okay.' But knowing the source is the prerequisite, and we just got good at it.