The first line is wrong.
It’s not that machines “can’t develop shared understanding” it is that machines “do not yet” apply Bayesian probabilities to developing a context sensitive model of the conversational context. That will come. It is not a huge step from where they are currently.
As I have stated many times before, the major key to understanding human cognition is getting that we do not experience reality directly, rather we experience a model of reality created by our brains from a mix of past and present experience, and the distinctions and concepts that seem contextually relevant.
When our machines start producing similar contextual mapping, the accuracy of understanding will improve hugely.
Context includes such things as the individuals present, the physical context, the topics and levels of conversation, the shared history between individuals.
Machines don’t do that yet.
There is no conceptual barrier to them being able to do it in the future.
I can think of three very different technologies that could make that happen.