[schoen]

Maybe a more intrusive setting would be something like a polygraph. The subject would be first tested to ensure that the training worked to a high accuracy level (like the polygraph baseline measurement). You could tell to some extent whether the subject was cooperating with this training because you could see the resulting accuracy rates, and maybe continue the training until they improved, or else declare the subject to have failed to cooperate with the test.

Probably, as with an present-day polygraph, people could specifically train themselves to produce inaccurate results, but also presumably most people wouldn't have done so.

On the other hand, I don't know what level of language this system operates at so I don't know if you're detecting phonemes, words, phoneme features (like +voice/-voice), or what. This seems significant for understanding how usable it would be for "mind reading" if the subject's thoughts included many words that hadn't been individually trained. If you're just detecting conditions that you've already trained for, it doesn't seem like there's really anything new here (from the mind-reading point of view, as opposed to the UI point of view) compared to existing biofeedback or polygraph systems because you're mainly just detecting when a known physiological state is or isn't re-elicited, which is something the earlier systems could already do.

Edit: other commenters said they can only distinguish about 20 words, so presumably the technology has a long way to go before it would be relevant to semi-adversarial mind reading.