It's interesting but can't work outside of a lab. The example they gave was watermarking a picture of a cow with a purse. If every cow picture has a hidden purse watermark, an AI will be trained into categorizing a cow as a purse.
But to make that work in the real world would require every artist and especially stock photo sites to agree to watermark every cow with a purse. If everyone doesn't pick a consistent watermark of a purse for a cow, then it becomes noise that is trained out. Just like training a to identify a cow sometimes has a farmhouse in the picture, other times grass, other times birds. It learns cow because that's the consistent part. Without a purse watermarked into the majority of every cow photo everywhere, the ai will learn cow.