As is typical, this science reporting isn’t great. It’s not only that AI can do it effectively, but that it can do it at scale. To quote the paper:

“Despite these models achieving near-expert human performance, they come at a fraction of the cost, requiring 100× less financial and 240× lower time investment than human labelers—making such privacy violations at scale possible for the first time.”

They also demonstrate how interacting with an AI model can quickly extract more private info without looking like it is. A game of 20 questions, except you don’t realize you’re playing.