Why is AI safety so much harder than Isaac Asimov's "Three Laws of Robotics"?

Question

I understand that AI researchers are trying to create AI designs that allow for desired behavior without undesirable side-effects. A classic example of an attempt is Isaac Asimov's Three Laws of Robotics. This idea seems to have been debunked due to its vague phrasing. In particular, I realize that the exact 3 laws from the stories can't work, but surely there is a set of more robust ones that can limit an AI to good behavior in the same way that people are restricted by law.

Why have AI researchers not accepted an idea like the following (just making the laws more specific):

What if the UN voted for the country with the fairest laws and all utility functions have 2/3 of their points made up of not breaking any of those laws. If the ai has a question, it could look at court precedent just like a judge would or ask humans (with two-thirds of its points on the line, it should be pretty cautious).

People have been looking for loopholes in law for thousands of years and there may not be any catastrophic ones left. (It certainly wouldn't be violent)

I must be missing something if this is still an open problem.

Answer 1

The reason this is hard is because it is not trivial to understand what a law means. Many humans still have a hard time understanding laws and thus we have millions of judges and lawyers who study years to be able to even debate whether a law was broken at all.

More generally to AI, the problem of understanding laws is a byproduct of the bigger problem that AIs today do not yet have mechanisms to turn input (text, audio, other sensor data) into concepts. This is evident in even the most advanced conversation systems which struggle with dialog which is off of a particular set of subjects.

However, there are promising results with probabilistic logic programming languages that attempt to combine first-order logic with statistical inference ($\text{LP}^{\text{MLN}}$ is an example). These languages could provide, in theory, a way for future AI-lawyers (or whatever one should call them) to encode exactly what laws mean. Using this set or encoded laws, a system with a rich enough representation could determine if an action would brake a law or not. This sounds great but the "rich enough representation" is beyond what even today's best deep learning systems (or any other architecture) can do. For comparison, today's best deep learning systems don't even have good object permanence, something humans have around 4-7 months old. Asking those deep learning systems to also be able to represent the complexities of laws is profoundly outside of current AI capabilities.

Another promising area is only including text data and having artificial judges. This wouldn't immediately mean that AIs would obey laws but it is the start of AIs understanding the word of law.