Do you think they are using AI as they observe? Is some of this AI passive? Obviously they wouldn't be able to spy on everyone, or else they would need a lot of researchers haha
From the article, it looks like they have observed that the chat data is used for cross referencing image hashes to text, but extending it to incorporate and leverage representations learned via ML algorithms (supervised, or self-supervised) should be quite straightforward.
They probably use something like NarusInsight, which according to Wikipedia, can do real-time semantic analysis.