3 ms·

There are some characteristics, shaped by evolution, that are typically human. For instance, humans are social. That doesn't mean that you can derive a formal

There are some characteristics, shaped by evolution, that are typically human. For instance, humans are social. That doesn't mean that you can derive a formal system of ethics that is "true", but there are clearly some trends. Trends that an AI, that have not evolved as us, don't need to share. I don't think is a sterile debate.

AndrewKemendo 8y ago
For instance, humans are social As are honeybees. I fail to see how this (or any) feature is distinguishing when describing how one should build an ethical framework into an engineered system. If however there was a more nuanced pro-social univerally attributable aspect that could be described and modeled, then I think that would be a helpful addition. I don't think is a sterile debate. It's anything but sterile. However what is sterile is the idea that there exists a desirable ethical system for AGI that is based on human ethics.
- RobertoG 8y ago
  >>"I fail to see how this (or any) feature is distinguishing when describing how one should build an ethical framework into an engineered system. " My point is that human ethics is grounded in human values and those are grounded in our evolutionary history. As we share an evolutionary history it would be not strange that there are a lot of values that we share. Maybe they are difficult to see because they appear to us as obviously true, but this is not going to be the case with machines. Presumably, we want machines that care about the same things that us.
  - AndrewKemendo 8y ago
    Maybe they are difficult to see because they appear to us as obviously true I argue they are difficult to see because they don't exist. For example "Don't Kill" would be the obvious values that we share - until it's not a value we share. Maybe it's a value we ought to share, but as the dozens of examples in philosophy show (trolley problem etc...) in fact it's not something that we can hard code. Ok, so maybe it's "it's ok to kill in the following circumstances..." in which case a brief history of humanity would show massive changes in what those circumstances are over the last 10,000 years - so that's a bad candidate for a "value" set as well. Presumably, we want machines that care about the same things that us. This is the going argument and has been since the concept of "friendliness" in AI came about. Trouble is, we can't seem to pin down the things we as a human race collectively "care about" to the granularity that would be required to write machine instruction around it. I'm arguing there aren't any collective values that everyone would agree upon which would inform such a discussion.

AndrewKemendo 8y ago

For instance, humans are social As are honeybees. I fail to see how this (or any) feature is distinguishing when describing how one should build an ethical framework into an engineered system. If however there was a more nuanced pro-social univerally attributable aspect that could be described and modeled, then I think that would be a helpful addition. I don't think is a sterile debate. It's anything but sterile. However what is sterile is the idea that there exists a desirable ethical system for AGI that is based on human ethics.

RobertoG 8y ago
>>"I fail to see how this (or any) feature is distinguishing when describing how one should build an ethical framework into an engineered system. " My point is that human ethics is grounded in human values and those are grounded in our evolutionary history. As we share an evolutionary history it would be not strange that there are a lot of values that we share. Maybe they are difficult to see because they appear to us as obviously true, but this is not going to be the case with machines. Presumably, we want machines that care about the same things that us.
- AndrewKemendo 8y ago
  Maybe they are difficult to see because they appear to us as obviously true I argue they are difficult to see because they don't exist. For example "Don't Kill" would be the obvious values that we share - until it's not a value we share. Maybe it's a value we ought to share, but as the dozens of examples in philosophy show (trolley problem etc...) in fact it's not something that we can hard code. Ok, so maybe it's "it's ok to kill in the following circumstances..." in which case a brief history of humanity would show massive changes in what those circumstances are over the last 10,000 years - so that's a bad candidate for a "value" set as well. Presumably, we want machines that care about the same things that us. This is the going argument and has been since the concept of "friendliness" in AI came about. Trouble is, we can't seem to pin down the things we as a human race collectively "care about" to the granularity that would be required to write machine instruction around it. I'm arguing there aren't any collective values that everyone would agree upon which would inform such a discussion.