But it is deterministic. It might be far too complicated to be predicted in practice but in principle code will always lead to completely predictable outcomes. Some think that this is true of consciousness as well to be fair but I don't think free will is an illusion.
Only if you don't allow feedback after training, and almost no one does this anymore. That means these systems are inherently NON-deterministic, because each run through a neural network path may cause a change in connections, and those changes are driven by insertion of a random number to pick the changes.
Now the goal of using AI is to produce "generally convergent" behavior, but the problems that most people are trying to solve with AI include far more permutations of cases than initial training. That means that any particular input may throw the current state of the network into a state it hasn't seen before and force it to "try something new" no matter how small.
This is a gross oversimplification of what actually happens, but it's pretty accurate.
Even VonNeumann architectures can mimic this with full non-determinism, but we're starting to move the "hardware neural nets" and rediscovering the joys of analog architectures, with quantum mechanics thrown in not only for parallelism but for "true-random" non-deterministic sources.
It's all pretty incredible, and if you haven't been watching it carefully, you'd be amazed at how different AI is than just 5 or 10 years ago.
Because it's just code. No amount of computation can generate consciousness. If it could then all physical processes would have to be conscious, at least a small degree. The processes of the brain can't generate consciousness any more than computer code can, instead it's consciousness that controls the physical processes.
Consciousness obviously controls our bodies but also our brains. If you make yourself feel a certain emotion then your brain will react to that, we are in control of it rather than it being in control of us. Like all computers, our brains can only follow instructions. Somehow our consciousness is able affect physical reality, specifically our brains, probably through a quantum mechanical process like taking advantage of the uncertainty principle.
So, by removing your premise that it's all completely deterministic, that makes the question of "generating consciousness" not at all clear.
The whole point in AI of using neural nets is that neurons are the fundamental building block of brains, and while those nets don't exactly work the same way--yet, watch this space!--they do provide a structure for dynamic self modification that we can see wetware brains doing. We don't understand either very well, but there's no justification for saying that what's being worked on cannot, under any circumstances produce what some people mean by "consciousness."
And of course the fact that we can't really define "consciousness" in any precise way, means that "replicating it" is more a matter of artistic interpretation than analysis of facts.
In my opinion.
Ay, there's the rub,