I'm saying that if you can detect a neural signal which is clean enough to show detectable, repeatable changes when someone focuses on a particular word, then there are algorithms which can be used to classify that signal as corresponding to one of several pre-defined words. Perhaps 50 or so words. I'd be surprised if they could do well on more than that; but a trained soldier could communicate a host of concepts to his men with even that much of a vocabulary.
It's not mind-reading, it's just pattern recognition. Train an engine on 100 inputs corresponding to "yes" or "no", and there's a good chance that given a 101st signal, you can correctly classify it as yes or no. You won't get it right all the time, but you'll do better than chance----and once you're better than chance, there are a host of algorithms you can employ at a higher-level to improve your results.
Depending on whether the classification algorithm is generative or discriminative, you might also be able to generate "typical" signals which correspond to a particular word. No telling whether they'd really be meaningful or not, though. If they were that *might* be a minor step towards mind control, but we're well away from that being a real problem I'll bet.
Now, the real problem here is (a) finding the particular signal which will provide sufficient discrimination between words, and (b) detecting that signal with a high enough SNR to be useful with a low-power portable device. Given that DARPA is funding this at all, I'm guessing someone has already demonstrated (a).
