I agree that simulation is the key to a concious AI system. If or when we ever succeed in simulating a human mind to a close enough degree of comparability, it is almost a given that the system will be self concious.
There are some problems simulating the human brain that would also have to be addressed even once we can create a working system, such as the AI being a bit of a blank slate, like a infant or a coma patient.
I see the whole process as having to follow a path similar to this:
1. A breakthrough in computing power, something capable of simulating very accurately small areas of space, this means perfectly representing ridiculously complicated chemical reactions and some natural laws.
2. Succeeding in creating a software environment to execute these simulations within.
3. A breakthrough in mapping an existing person, some sort of scan that creates a mathematically provable perfect (or close enough) representation of an area in space. Like some humans mind or possibly their entire body until the subject of the scan can be simplified on the computer. Sort of like taking a photo and then cropping off the body. The above simulation environment may be what is used to provide the simulated inputs and outputs to the head, like the CNS and cardiovascular system. Not to mention the inputs to the eyes and other senses. Sort of like a virtual head in a jar.
4. So far we would have a conscious system, but it would be a copy of a pre-existing being. The next step would to be to somehow, ethically, re-write this being. This would provide a learning challenge with the goal of simplifying and modularising the human brain. Such as hacking language areas, input nerves, the reliance on virtual blood and sustenance and most importantly the memory.
The final product of this important stage is the most simple and easily tweak-able simulation of the human brain that could be used by all researchers and eventually commercial applications. If all these virtual brains are the same or comparable, this isolates the memory as a way to load in or edit what is essentially... people. The creepiest analogy may be the best, they will be like swappable save game files, executing in virtual machines (the hacked brains) that operate within another virtual machine (dare I say it, a super-simple matrix of sorts)
5. We may never reach anywhere near this far along the process due mainly to ethical reasons that cannot be overcome with mere ingenuity. But if we do, the next step is compressing all this down further and further until we have the most simple possible (perhaps provable somehow) implementation of a mind that does not require all the layers of virtualization.
God it's easy to get caught up in this stuff. I hold this prediction on my fingertips in hopes that any developments may blow it away so I can re-evaluate and make a new one.