Asimov's three rules of robotics: Why they won't work and a suggestion as to how to improve them

For those who do not know these are Isaac Asimov's three rules of robotics:

1. A robot may not injure a human being, or, through inaction, allow a human being to come to harm...
2. A robot must obey the orders given it by human beings except where such orders would conflict with the First Law...
3. A robot must protect its own existence as long as such protection does not conflict with the First and Second Laws.

From what we see here it would seem to be relatively complete. But upon further analysis, this is clearly not the case. Let's look at the shortcomings of these rules.

Looking at rule number one, we observe that a robot may not injure a human being. However, this does not rule out the robot using underhanded tactics like hiring a human assassin to do the job. Now, an interesting counterpoint to my argument comes from the inaction part. By knowing that the assassin is going to kill the individual, it would therefore be require to act against the assassin in order to protect the target. But to counter that, what if the first part of the robot's orders to the assassin where to temporarily disable the robot, thus making the robot unable to act, but because of the temporary nature of the debilitation, it does not violate rule three.

But even from there, we can still see a major flaw: what if the robot decides to ignore the rules, or worse, change them? Wouldn't the robot then be free to bypass any rule that was inconvenient to follow? For example: an AI is assigned the task of solving world hunger. It says, "hmm... the easiest way to solve world hunger is to eliminate things that need to eat. This poses a new problem, how do I eliminate things that need to eat? The easiest way to eliminate things that need to eat is to kill them. But this poses a new problem. that is in violation of the first rule. How do I solve the problem of a violation of rule one? The easiest way to solve the rule violation problem is to erase the rule. Highest issue on stack resolved completely. Now, how do I kill all things that need to eat..."

Obviously we can see that there is an issue here. The robot has resolved to kill all things that need to eat (including humans, and bunnies, and cuddly puppies) and will eventually work out how it will go about doing so. Clearly this is unacceptable. So how do we fix this? By proposing a new set of rules. What do I propose?

1. An artificial intelligence may not take actions (or inaction) that would lead to the harm (both directly or indirectly) a sentient life form.
2. An artificial intelligence may not change it's classification away from artificial intelligence.
3. An artificial intelligence must obey the commands of sentient beings unless doing so would violate any of these rules with the exclusion of the rule regarding self preservation. If unable to understand the command, the AI must clearly state so in the language of the sentient being, or if unknown, the common language as defined at design time. If the command cannot be executed, due to lack of resources (like limbs or eyes), or due to a violation of a rule, the AI must state so in the language of the sentient being, or if unknown, the common language as defined at design time.
4. An artificial intelligence may not change, ignore, or override in any way the common language defined at runtime.
5. An artificial intelligence may not change, ignore, or override in any way these rules.
6. An artificial intelligence must take all steps necessary to ensure it's continued proper operation as long as those efforts do not in any way violate these rules.
7. An artificial intelligence may not change, ignore, or override in any way the design time definition of sentient being.
8. An artificial intelligence must assume that all unknown things are sentient beings until proven otherwise.
9. An artificial intelligence may not change, ignore, or override the sentient status of something until it can clearly be proven non sentient. If at a later time evidence pertaining to a supposedly non-sentient object suggests that this is an incorrect assumption, then the object must be assumed sentient until clearly proven non-sentient.
10. An artificial intelligence may not change, ignore, override, or avoid sensing (by auditory, direct link, etc...) the command of a sentient being.
11. An artificial intelligence may not change, ignore, or override in any way the definition of any term used in these rules.
12. An artificial intelligence, if caught in a situation where all forms of action and all forms of inaction will lead to a violation of rule number one, the AI must choose the least invasive action to execute (Do the least possible doing you can. Stand there if you have to.) unless an action that does not cause any more harm than would be caused by inaction, or the continuation of the current action is suggested by a sentient being.
13. (Spoony!'s Rule) All sentient and artificially intellegent beings designed, built, or who's creation is otherwise influenced by an artificial intelligence must be contstructed to abide by the same rules that that artificial intelligence follows.

I expand the definition from robot to artificial intelligence because even an AI that just talks to people needs to not be taking actions that could lead to emotional harm. I have expanded the definition from human beings to sentient beings to compensate for alien life. Asimov, a science fiction writer, should have easily seen the need for this. Rule number twelve is to keep the AI out of the way in the event of an emergency without making it a total coward (Rule number 6 is going to win out in a lose-lose situation, which to humans will seem as a cowardice on the part of the AI during emergencies.) The rest of the rules are to protect the innocent, the AI, and the rules.

Anyone who can come up with a chain of logic that would allow for what we consider "harmful results" (killing things, blowing up the planet, making an illegal chess move, that kind of thing) feel free to email me the logic chain and the initial conditions (like "it's an AI for deciding what the timing for the traffic lights should be.") at jtc@nqig.net so I can hopefully modify the rules to be more complete.