Artificial Intelligence Stack Exchange is a question and answer site for people interested in conceptual questions about life and challenges in a world where "cognitive" functions can be mimicked in purely digital environment. It only takes a minute to sign up.
Sign up to join this community
Would it be possible to put Asimov's three Laws of Robotics into an AI?
The three laws are:
A robot (or, more accurately, an AI) cannot harm a human being, or through inaction allow a human being to be harmed1
A robot must listen to instructions given to it by a human, as long as that does not conflict with the first law.
A robot must protect its own existence, if that does not conflict with the first two laws.
The most challenging part is this section of the first law:
or through inaction allow a human being to be harmed
Humans manage to injure themselves unintentionally in all kinds of ways all the time. A robot strictly following that law would have to spend all its time saving people from their own clumsiness and would probably never get any useful work done. An AI unable to physically move wouldn't have to run around, but it would still have to think of ways to stop all accidents it could imagine.
Anyway, fully implementing those laws would require very advanced recognition and cognition. (How do you know that industrial machine over there is about to let off a cloud of burning hot steam onto that child who wandered into the factory?) Figuring out whether a human would end up harmed after a given action through some sequence of events becomes an exceptionally challenging problem very quickly.
Defining "harm" and in particular, "allowing harm via inaction" in any meaningful way would be difficult. For example, should robots spend all their time flying around attempting to prevent humans from inhaling passive smoke or petrol fumes?
In addition, the interpretation of 'conflict' (in either rule 2 or 3) is completely open-ended. Resolving such conflicts seems to me to be "AI complete" in general.
Humans have quite good mechanisms (both behavioral and social) for interacting in a complex world (mostly) without harming one another, but these are perhaps not so easily codified. The complex set of legal rules that sit on top of this (polution regulations etc) are the ones that we could most easily program, but they are really quite specialised relative to the underlying physiological and social 'rules'.
EDIT: From other comments, it seems worth distinguishing between 'all possible harm' and 'all the kinds of harm that humans routinely anticipate'. There seems to be consensus that 'all possible harm' is a non-starter, which still leaves the hard (IMO, AI-complete) task of equaling human ability to predict harm.
Even if we can do that, if we are to treat as actual laws, then we would still need a formal mechanism for conflict resolution (e.g. "Robot, I will commit suicide unless you punch that man").
I think this is almost a trick question in a sense. Let me explain:
For law 1, any AI would abide by the first rule unless it was deliberately created to be malevolent, in that the AI it would understand harm was imminent but do nothing about or would actively attempt to harm. Any 'reasonable' AI would (try its best to) prevent any harm it understood, but couldn't react to imminent harm 'outside it's knowledge', thus satisfying law 1. Any AI that 'tries its best' to prevent harm works here.
For law 2, it is simply a matter of design. If one can design an AI capable of parsing and understanding the entirety of human language (beyond just speech), just program it to act accordingly, mindful of the first law. Thus, I think we can develop an AI that will obey every command it understands but getting it to understand anything and everything I believe is impossible.
For law 3, it rides in the same vein as law 1.
In conclusion, I think there is no philosophical problem with implementing such an AI, but that the actual design of such an AI is fundamentally impossible (understanding all possible harms, and all possible commands).
The paper The First Law of Robotics (a call to arms) (AAAI-94), by Weld and Etzioni, discusses the first Asimov's law, some technical issues it gives rise to (some of them are already mentioned in the other answers), and how they could be addressed (they propose a simplistic way to formalize the first law, but they don't claim it is the right way to do it). You should read it for more details.
