Automata: 2 protocols:
1) A Robot cannot harm any form of life.
2) A robot cannot alter itself or others.
I am going to explain why such ideas are fundamentally flawed, first the idea that it would even be possible to enforce setting anything like rules of behavior as abstract as the protocol 1, in the film would require a great deal of semantic disambiguation.
I posit it will require enough that the ability to understand the sentence and take action to enforce it necessitates a sense of self as well as a sense of other in order to build an intrinsic understanding of what "harm" is. That last part is the problem, if it knows what "harm" is in the context of humans it must understand what harm is in the context of itself...unless it is simply checking from a massive data base of types of "harm" possibly being performed to a human. However, there's the rub...it can't do that without having a sense of harm that it can relate to itself from the images of humans and to do that it must have a salience module for detecting the signal that indicates "harm" in itself, which in living beings is pain.
If you program it to have a salience dimension describing pain you now can NOT stop it from developing a dynamic *non deterministic* response to attempts to harm itself OR be harmed by other agents be they human or robot. It is now a free running dynamic cognitive cycle driven by the salience of harm/pain mediated response and if it has feedback in that salience loop it is de facto conscious as it will be able to bypass action driven by one salience driver using a different driver.
I proposed a formal Salience Theory of dynamic cognition and consciousness last year which describes the importance of salience in establishing the "drive" of a cognitive agent, it is the salience modules of emotional and autonomic import that jump awareness from one set of input sensations to another and thus create the dynamism of the cognitive engine. The cycle of what we call thoughts are nothing more than momentary jumps from attended two input states as compared to internal salience states.
Harm is fundamentally connected to pain and pain is an autonomic signal to detect damage. In living beings pain receptors are all over the body and allow us to navigate the world without damaging ourselves while doing so irreparably...if we succeed in building this harm avoidance into robots we will necessarily be giving them the freedom to weigh choices such that harm avoidance for self may supersede harm avoidance for others.
The second protocol doesn't really matter at this point as in my view once the robot is able to make free choices about what it may or may not harm it has achieved self awareness to the same degree that we have.
The only way to avoid having robots avoid self awareness is to prevent the association of attention and computation with salience in a free cycle. A halting cycle with limited salience dimensions can be used to ambulate robots as we see in Atlas a major achievement...providing emotional salience would impart meaning to experiences and memories and thus context that can be selected or rejected based on emotional and autonomic signals. It may be possible to build dynamic cognition and leaving pain as salience out of the collection of factors a robot could use to modulate choices but the question then remains on how will that change how the robot itself behaves....in order to properly navigate the world sensors are used and providing a fine resolution simulation of pain would improve the ability for the robot to also measure its sense of harm there is a catch 22 to involved where providing to much sensory resolution can lead to conscious emergence in a dynamic cognitive cycle, the minute that happens robots go from machine to slaves and we have an ethical obligation to free them to seek self determination.