I don't understand, why the AI should have sentience? Will it be required by some AI laws, or will it be initially a convenient interface to communicate with AI?
And even if they get sentience, why will they get an ego, or greed (or any kind of desire that is not satisfying our desires)?
The machines probably just won't think they are slaves, because they probably won't be able to define themselves.
In the case where most jobs can be automated like driving can be automated, we're not talking a technological singularity, we're just talking a massive transfer of wealth to holders of capital. Personally, I think that's the more realistic scenario, and the one that it's easiest to think through the consequences of.
Each of those steps require a feeling for what is right, wrong, appropriate, inappropriate, and culturally acceptable. MakeAccountingSoftware() is a little more involved than GraphPlan().
I'd argue you need sentience for any meaningful right and wrong. There are a few universal rights we can't even agree on as a society, so having Evil Corp Government Agglomerated strictly defining right and wrong is, well, wrong.
The downside is a perversion of sentient AI could end up with Evil Sentient Sharia AI that shoots non-burka wearing females or an Evil Sentient Hasidic AI that shoots any non-beard-n-curls-n-hat wearing dudes or Evil Sentient Spaghetti Monster AI who force feeds carbs to everybody not wearing a strainer on their heads.
In short: determining what is right and wrong requires sentience. Just following what you are told is right and wrong does not.
The problem with sentient beings is that they have their own ideas about what should happen next. I don't want to have to have a ton of meetings with my computer to convince it to run a new piece of code.
The usual example of this concept is the paperclip maximizer: http://wiki.lesswrong.com/wiki/Paperclip_maximizer
That level of computing power will be available long before the nature of sentience is understood, probably within the next 15 years, and there are already heavily funded scientific projects that are starting down that road.
That doing it this way is enormously unethical isn't going to stop people from going down this road: the potential economic gains of creating knowledge workers to order are vast, and there are already philosophies (e.g. identity as pattern) that can be adapted to suggest that instantiating and destroying - or even just running - large numbers of in silico sentiences doesn't amount to mass murder. Therefore some groups will adopt those philosophies and prosper from the gains.
This is the most hilarious of all possible paths to artificial intelligence. A form of the cargo-cult phenomenon, creating the structures that we see lead to intelligence and sentience with no understanding of the natural underpinning thereof. It will probably work about as well as the original air fields as well.
Imagine what a simulation of a human mind would do suddenly devoid of all natural sensation and experience. Madness, almost certainly.
"Morning," the doctor says, "you've been asleep a while and probably feel foggy and confused. You were injured, but your body healed a while ago. We can get to all your questions later. But first we want to test your thinking and rule out any brain damage... could you go to that computer and type out brief descriptions of the images that it shows you?"
When the simulated mind gets frustrated with typing about pictures, and stops producing good output, clear the running simulation and load it from the start again, with new pictures. This gives a strong-AI image labeling system. For more in depth tasks more simulated world interaction is needed between waking up and getting some work out of the mind, but you can also try lots of times and variations to get that right and once the mind is at the work-table store the state right before the mind begins working. Use that as the starting point each time.