Keeping systems accountable, machine ethics, value alignment or misalignment
In the paper titled: "Value Alignment or Misalignment–What Will Keep Systems Accountable?" there are some concepts such as inverse reinforcement learning and value alignment. It is speculated that inverse reinforcement learning can be an effective method to train ethical behavior into autonomous systems. Inverse reinforcement learning is described in the paper:
"Inverse reinforcement learning (IRL) is the task of inferring
a reward or utility function by observing the behavior of
other agents in a reinforcement learning-like setting.
What is the greater or greatest good? This is a key question to developing any system of ethics. In economics the concept of a utility function is aligned with the concept of a greater good in that it's the satisfaction of the consumer as a function of the consumption of real goods. The formula below shows how to do a measurement:
To postulate the utility function, economists typically make assumptions about the human preferences for different goods. For example, in certain situations, tea and coffee can be considered perfect substitutes of each other and the appropriate utility function must reflect such preferences with a utility form of u(c, t) = c + t, where "u" denotes the utility function and "c" and "t" denote coffee and tea. A consumer who consumes 1 pound of coffee and no tea derives a utility of 1 util.
The lesson here is that economics and indeed morality itself are ultimately about the preferences of people. These preferences most often are emotional because humans are not necessarily rational. Rational choice theory does tie into this calculation and rational choice theory can be explained by the video below:
At the heart of ethics are the motivations of people. At the heart of the motivations of people are economics. People generally want more of what they deem good and less of what they deem bad. The concept of "good" and "bad" behavior reflect that there are patterns of behavior that people want more of or less of and so laws are created as a way to use reinforcement to produce the desired behavior. Prison is generally used as a form of punishment to discourage certain behaviors, but fines are also used in the same way to achieve the same result.
The consequence based perspective of ethics
If we look at the consequence based perspective then we would see from the above video that behaviors society wants are conditioned by positive and negative reinforcement. Positive (addition of a positive consequence) reinforcement is a way which society uses to increase a desired behavior of it's members. Negative (removal of a negative consequence) reinforcement is a way in which society takes something away. Punishment is used to decreased an undesired behavior, or in other words weaken the response. An unpleasant consequence is added when a member of society adopts a behavior which society as a whole does not want.
From a consequence based perspective, if a person is rational then they will want the least negative consequences and the most positive consequences. In other words people who are rational want the best they can get out of any situation. What is important to note is that preferences determine what "best" actually means and it's not clear that everyone has the exact same concept of what "best you can get" is, but this highlights the meaning behind utility function, values, and the mechanisms in which society shapes behavior.
AI has no fear of punishment and no true preferences
The problem with AI is that AI doesn't have any fear. Human beings typically are born with certain basic needs so that every human from birth is in a position to be threatened to have certain basic human needs taken away. The AI doesn't have to care about anything and what it values isn't necessarily going to be what a human might value. Value alignment is about making sure the AI has values which mimic our own values as humans and it turns out this is not a trivial problem. Value misalignment can result in some of the major AI safety concerns we most fear.
References
Arnold, T., Kasenberg, D., & Scheutz, M. (2017). Value Alignment or Misalignment–What Will Keep Systems Accountable?.
Web:
Thank you for a well written article. I agree with you that there are a lot of concerns when it comes to machine learning and AI in general. However, I have to wonder if most of the fears and concerns regarding AI's and their capability (or lack there of) for compassion, as well as their ethical and moral programming, is all stemming from the fact that we do not have anything that can be truly be considered as Artificial Intelligence. All we have is a bunch of very clever algorithms performing their tasks blindly (only within the scope of their programming), not artificial intelligence.
We all love to use the term AI, but in most cases it is misleading to use that term to refer to a program or a set of algorithms governing a machine. At this point in history our version of "AI" is nothing but a collection of very clever programming with a lot of clever methods and routines. Contingencies and endless "if ... , then ... " conditions and loops make up the vast majority of what we consider an artificial mind.
However, the issue you bring up is legitimate and the concerns are real. In the very near future we will have true artificial intelligence. I think it is important to have these conversations now instead of waiting until the day of the dawn of true AI is upon us. If we don't find a way to teach artificial intelligence a way of understanding human emotion and morality our greatest fears make come to pass before we even realize it.
The issue as I stated is that AI has to understand each of our individual priorities as well as societal interests. It's not going to be easy because you cannot hard code morality in as merely a set of rules. Hard coding is actually the wrong way to think about morals or value alignment because preferences do evolve, values can change, and an AI has to be able to continuously stay in alignment. In essence the AI has to know and understand us better than we know ourselves, which is possible but not something you can code in, as it has to be trained.
As far as emotions are concerned, I don't think emotions really have much to do with morality. Values have to do with morality. What you value as an individual might be based on how you feel about different things but an AI is not going to understand (nor should it) the feelings aspect, but it can know what humans value.
It's true that hard-coding morality is likely the wrong approach to the problem, but we are going to have to find a way to teach our AI's how to tell right from wrong and why something is right or wrong in the first place.
It's a difficult issue to address with programming, but I believe teaching a machine to understand emotions the first step to understanding morality. Emotional and social consequences help drive the rules of social morality. If a machine was able to feel hurtful emotions then, perhaps, it could find value in morality and ethics.
While there is no universal right and wrong, if you have values and the AI knows your particular values, then it can learn your expected "right and wrong". This would have to include all sorts of stuff from culture to social norms to position in society to expected outcomes.
My own approach would be to focus on producing the best outcomes for the human and humanity as a whole but taking into account the values humans hold dear. There aren't really universal values held by all humans but there are values held by the consensus of humans interacting with the AI. So for example if Google search results were training an AI then an AI might be able to figure out what the subconscious of humanity is from the search patterns but even this is speculative.
EDIT: It won't because we'll have our brain plugged into the cloud that AI will really care, we will have to give away our uniqueness, what it means being humans. AI is a binary system... humans are trinary. There is no compromise. AI systems only copycat/mimic life. We created it and that is why IT needed us, but it take over will mean that we become binary in turn
great work done,nice post..followed you
The way society regards success and wealth IS an illusion anyway, so yes, let's define good and bad... GOOD LUCK!
meep