If these networks were only being used to model other systems, like protein folding or something like that, people wouldn't be mistaking them for anything other than clever prediction engines.

Since they're modelling our own languages, people get spooked and start bringing up theory of mind.