Pretty soon they will be able to run all our information through a language model.

Models could identify "suspicious" behavior and generate plausible theories to fit the desired witch hunt, based on what they read. Scary.

That is essentially what has been done since the FBI Carnivore software circa 1997. The software gets more complicated but the goal is the same. I assume that the drive for better AI models is in part driven by intelligence agency needs to categorize mountains of data to build profiles of individuals.

The eventuality is of course a dystopian thought police situation, where simply thinking or even Googling something suggesting you were even contemplating something illegal will be grounds to charge you with the crime.