When I see "horse poop" the thing I'm imagining and hearing in my head is not the individual letters as they appear on the screen at all. I'm hearing it in an internal monologue and I'm imagining and image and a smell.
In other words, the output generates wildly different input so 'subtracting it out' won't really work.
We've all seen people who are new to keyboards, mice and even GUI's attempt to navigate a computer. It's clunky, slow, there's a lot of noise in the movements and erred clicks made.
I think it was here on HN I saw a great write-up once about leaving the gap in UX—that there is a threshold where the use just simply has to learn something in order to effectively use the device.
I don't know how the control interface for thought could look in the end in so few words, but I'm reasonably confident it could be brought down to a certain level that is more than attainable for the average user to meet.
The same procedure would be more difficult to implement for thought-based interfaces, though, because you need to predict the brain's reactions to filter the signal. Maybe you could instead use a non-verbal thought to activate the command interface, so that it doesn't get triggered accidentally.