The brain machine interface concept seems very useful. My question is the AI part. Aside from the machine learning likely needed to decode brain signals meaningfully at all, why would we want to hook up something with any resemblance to current AI to a brain directly?

The current AI stuff can easily be described as a breakthrough in natural language interfaces. A field engineers have been working on for many years. It easy to imagine that the methods used to develop current AI could be used for different types of interfaces we've been stuck on.