Check out https://www.nature.com/articles/s41586-019-1119-1
And yes it is definitely orders of magnitude harder than controlling a joystick
With a few decades (years?) of refinement, this technology absolutely has the potential to capture private thoughts.
Sure, it may require an implant, but just like we all carry tracking beacons that have become indispensable to daily life today, in 20 years it may be “the done thing” to have an implant to control your home automation, etc. Or maybe the tech will improve to where it just needs a hat. And at that point, certain authorities will have no qualms about using it in interrogations, with or without “due process.”
It works because there is direct correlation between the speech-motor cortex and how the vocal articulators (larynx, tongue, etc) move, and how that combine to produce speech.
Abstract "thoughts", on the other hand, are not so straight-forward. For one, there's no central location in the cortex where the concept of "car", for example, lives. The distributed representation of abstract thoughts within the brain makes it orders of magnitude more difficult than decoding speech for one specific individual. Then add orders of magnitude of orders of magnitude to generalize that to different people.