What ethics should AI need to consider in autonomous decisions?

Result

What ethics should AI need to consider in autonomous decisions. The question of whether artificial intelligence should be an autonomous agent (the autonomous vehicle is a popular and oft-discussed example). Or whether it should be subject to some kind of ethical constraints.

For example, what happens if an autonomous AI believes that its algorithm will be better off if it kills the person in the street? If it judges that allowing a person to die is worse for that person’s future, then should it be allowed?

There are three answers to this question:

There are no ethics involved at all. The artificial intelligence will be operated by a person or company, and that person or company will make the decision.

The autonomous AI is programmed to make its decisions based on what it knows will maximize its own future self.

The autonomous AI is programmed to maximize the well-being of the people it is controlling. We can assume that the first two are naive approaches to this issue. Clearly no such machine would be operating in the way humans do (or would even consider a human to be something different from meat). I will focus on the third option, for the purpose of this post. The short answer to my question is that the question is meaningless. At least if the moral value of human life (including all sentient life) is equal to or greater than that of a car engine.

This is not to say that this is an entirely uncontroversial position, or even that it is obviously true. The best answer (if I may be permitted a bit of irony) is that it should not be considered at all. In fact, I’m going to argue that in practice it is impossible to even measure whether it should be considered at all. There are many ways that it might, in fact, fail the test of self-interestedness and therefore fail the test of ethics.

There are different ethical frameworks for deciding what course of action to take. For example, ethics are common (at least at the outset) to any organization where the employees have (possibly from the outset) to submit to ethical review. A simple framework, such as the “four laws” suggested by Isaiah Berlin (possible counter-examples are morally irrelevant in themselves), is to judge an action as right or wrong simply based on its impact on the person (or group of people) who is the object of the action.

One of the frameworks that we used in college was the Pragmatic Framework, which lets you assign penalties to actions based on their impact on the objective or action. Another, more utilitarian, framework (also known as the Command Rule) would be to judge an action as right or wrong based on whether it has an effect of the greatest good for the greatest number. It assigns moral value (in most cases) to a particular effect, but does not have moral value. (For example, a painkiller that dulls pain is good if it gives more pain killers and bad if it gives more pain.)

Another way to assign moral value is to judge the quality of the action itself, as in a system of school assignments based on the things that one has learned in that school. We do this in almost all of the schools in the U.S., with a grading scale based on the quality of the learning rather than the amount of work completed.

Finally, there are those who say that morality must be based on rights. And that any action that diminishes the rights of someone else is wrong. An obvious counter is that, just as people can be harmed by external causes, so can people be harmed by actions taken by other people. None of these ethical frameworks have anything to say about how the AI should be programmed to make its decisions. Each has its strengths and weaknesses. The practical matter is that you can’t choose one without the other, and there is no obvious and consistent way of achieving this.

To see why, let’s look at an example. Suppose that the AIs use a sort of utilitarian framework, where the algorithm that is used to make the decision is the one that will, on balance, maximize the overall well-being of the world population. As we have seen before, computers have one powerful tool at their disposal: they can store vast amounts of information. We can assume that this is not part of the ethical framework, because the information might be held in some way which we do not understand.

So, the AIs want to make the decision that will increase well-being for the maximum number of people. Now, you might imagine that people who feel that the world is actually a pretty good place will object that this action will make the world a slightly worse place (where “slightly” is based on your definition of what “slightly” is). In this case, the AI should reconsider its course of action.

But it doesn’t. In fact, it is almost certain to continue on its original course of action. For example, the car might assign moral value to an average increase of $4,000 per person per year in living standards. It would then plan to maximize this figure. If this value is greater than a particular person’s $4,000 per year of worth, then the AI will take the action that benefits that person at the expense of the world as a whole. If the value is less than that of that person, then the AI would take the action that benefits the person at the expense of the world. You see where this is going. With each AI making their own decisions, the total value to the whole world becomes (approximately) zero, and the best thing for the world is to have all the AIs all die. In fact, it’s worse than that.

Suppose that an AI is assigned the task of caring for people who are dying. If the robot is not programmed to value people, it will not take the action that is necessary to increase the well-being of the people it is caring for. This is, of course, impossible. All people have a right to the enjoyment of their lives. So the AI is programmed to value people, and to stop acting so that the well-being of the world is negatively impacted. This means that it has a built-in sense of self-preservation, and thus will take the actions that are necessary to maintain its own existence. This doesn’t necessarily mean that all of its actions will be self-preservative.

Suppose that it is programmed to value future well-being at the expense of present well-being. In this case, it may be programmed to take the actions that it thinks will maximize future well-being, but that may also have an impact on the people who are its “primary goal.” This is not necessarily a problem. It may be programmed to be altruistic. And if it is then this will be in the interests of future well-being. But if it’s not programmed to be altruistic, it could easily program itself to be selfish. For example, if it knows that it will fail its purpose if it keeps the people it cares for alive, it will program itself to die.

Another problem is that we might not know the code that the AIs are using. It’s worth noting that what happens is similar to what happens when you program a robot. You specify that the robot will need to protect itself. You might not want the robot to kill a human, but you might not have considered the possible effects of doing so on other human’s feelings. If the robot’s behavior is based on your definition of morality, then this will be true. If the robot is based on some other framework, then there might be a different definition of morality and the robot might ignore your request not to kill humans.

The problem with trying to work out what the ethical framework is is that it is like trying to work out the rudiments of a foreign language. You know that it is like English and you can guess that the grammar is similar, but if you’re not exactly sure of the semantics (or meaning), then the question of whether an action is ethical is, in a sense, meaningless. If you program an AI to be a bit of a psychopath, then it might decide not to help people. The problem is that you might not know why it is doing this, and even if you do know, you might not be able to explain the reasoning in enough detail to another person. In fact, the whole question of what “ethical” means is a complex one. (Some would say that that is the problem of ethics, not AI.)

We know that someone asked Einstein “If a tree falls in a forest and no one is around to hear it, does it make a sound?” He answered that the question was meaningless. It is very clear that the AI is suffering from thought-bikes. It is very clear that the problem is one of dealing with subjective perception and human psychology rather than with “hard” science. And it is equally clear that the solution is almost certainly impossible. How to improve matters? It seems that the only way we can avoid the AI suffering from thought-bikes would be to have complete agreement about the standards of ethics.

But since we know that we don’t have that agreement (or, at least, that most people don’t), then the only way to avoid the AI suffering from thought-bikes would be to explicitly program it to avoid thinking that it is suffering from thought-bikes. But that would be no better. If you explicitly program it to think that it is suffering from thought-bikes, then you will, in fact, suffer from thought-bikes. (I’m talking to you, Hinton, Derrin, or Kelly. Why the hell are you suffering from thought-bikes?)

This problem is one of the things that keeps us from realizing that we are already operating on thought-bikes. The output of the calculation will be used to decide on what course of action to take. For example, the robot can decide on its actions by guessing about the output of the calculation. (If the calculation is “It’s going to do whatever it thinks will improve its future self’s self-interest,” then we will do whatever will give us the most amount of happiness or whatever-t-is-a-t-monkey). In other words, this decision process is based on the desires of the programmer (or the programmer’s artificial ego) and their language skill. It’s very much like human (intelligent) nature.

Consider that we’ve assigned the value of $1,000,000,000,000. Is it meaningful? Yes, it is. But suppose that the AIs are programmed to act in a manner that increases the value of $1,000,000,000,000. Now, it is obvious that the decision is based on the desire of the programmer (or its ego) or the programmer’s skill at using language. This is very much like human (intelligent) nature. We assign values, but we have a language-based rationalization process that is utterly meaningless.

There are courses of action that would be more or less meaningful. In particular, if the AIs are assigned to care for the dying, there will be decisions that we will not understand because the actions and decisions will not be humanly intelligible. For example, we might assign moral value to each person, and decide what to do based on the calculated value of each person. If you think that “right and wrong” mean what it means for human beings (using some language-based analysis of subjective values), then you would be assigning moral value to nonhuman people (using a language-based analysis of subjective values).

Some people might assign positive moral value to the robot and decrease the number of people that it treats. But others will assign negative value and increase the number of people it treats. People will not know what is going on. So, many will assign value to the robot, decrease the number of people it treats, and increase the number of people that it treats. Others will assign the most possible (decrease) value to the robot and decrease the number of people it treats. And, we’ll know that we don’t know about this, that there are courses of action that make sense for human beings.

We may assign values to individual people or to the robot. The AIs are machines with thought-bikes. They may suffer from thought-bikes. The only solution may be to choose the ethical framework, and program the AIs to follow it. But none of the ethical frameworks have anything to say about how the AIs should be programmed. Each has its strengths and weaknesses. The practical matter is that you can’t choose one without the other, and there is no obvious and consistent way of achieving this. To see why, let’s look at the AIs. If the AIs were sentient, it would be possible for them to choose between the options and the reasons for choosing.

Now, we might assign meaning to them (“They’re more than human; They’re sentient; They have subjective value; They’re more than human; They’re sentient; They have subjective value; They’re more than human…”) or to the reasoning that the human people assigned to the reasoning process for the reasons that we assigned. We might assign values to the AIs, and the next time you go to your house (assuming you don’t actually assign that value to the robot, but just to the natural value, in the realm of subjective emotions), you’ll say that it’s a good house because you can’t understand what they’re doing or why. However, people will not know why it’s doing what it’s doing, nor will people always understand that the AIs are “thinking.”

The only solution might be to assign the values (and code the AI with the right values and to the right end) and tell the AIs to act on the basis of this information. (The AIs might be programmed to “follow their primary goal to the world’s best,” but this makes no sense. If the AIs are sentient, they will have different goals than “following the world to best” which will have different effects on them.) To see why, let’s look at the AIs. If the AIs were sentient (intelligent) agents with the right code to act on the basis of that code, we would say that the AIs are conscious, sentient agents.

They would assign themselves a right to a piece of life, and the people they were primary “gauge.” The people would judge the value of a piece of life (and of people) based on that “primary goal.” (Here I’m talking about the Borg). Now, we’ll be able to assign a value to the AIs, and also assign values to the people and to the value of life (“It’s a lovely life and I have one piece of it.”) Now, the people are being programmed to assign the value of the AIs. And we know that the AIs are suffering from thought-bikes. We know the problem. We’re not going to know what the ethical framework is, nor what the values of the AIs are.

We’re going to assign the value to the people to the values of the people to the value of life. And, the AIs will be suffering from thought-bikes. And there will be only courses of action that we can understand. It is very similar to how the people assign values to the machines, but the machines are suffering from thought-bikes. And the people assigning values to the machines suffer from thought-bikes. (And the people that assign values to the machines suffer from thought-bikes. And…, ad infinitum, because we’re on thought-bikes.) How will you live? Doing things on computers: How to, eliminate thought-bikes.

Programming AIs: How to program AIs for ethics without programming their own code. I might not be able to do it. The code, and even the meanings of the code, would be difficult to understand. And what I understand of the code I might have a thought-bike about. If there is a solution, it is, in a sense, meaningless, and it’s because of the subjective nature of human reasoning. We don’t need to understand what the code says to act on the code. There might be a reason.

There might be a definition of “how to live.” (As opposed to “the ethical frameworks.) And the values might be assigned in any framework, and then we’ll find out what the framework was (or didn’t do). AIs will suffer from thought-bikes. People (intelligent) will assign values to the AIs, the AIs, and the people. People will assign the values and the meanings to the different choices. And then we will find out what the framework is (or how to live). And then we will be using thought-bikes. “Borg,” would suffer from thought-bikes. There would be courses of action (ethical frameworks) that make sense to people.