How might we define morality or ethics such that we can program it into an Artificial Intelligence?

Question

With the speed at which AI is being developed, there is a pressing need to develop a clear system of morality that we might have an Artificial Intelligence operate from.

In my understanding, there are three dominant areas of morality and ethics:

Deontology - Kant

Utilitarianism - Bentham/Mill

Virtue Ethics - Aristotle

Have I missed any other possible moral/ethical codes that would be applicable to this discussion?

Which (if any) might be suitable to use as the default for AI morality?

Answer 1

As a starting point, a system of AI morality would have to be deontological, because for it to be something you can implement as a program, it would have to be a clear cut set of rules, as opposed to a utility measure.

Utilitarian ethics, even when applied to human situations, run into the difficulty of how to practically measure the utility of each situation. The problem of calculating such utilities would be impossible to implement in an AI system: It would be difficult to for an AI to have knowledge beforehand of all the variables that should go into such a calculation, and even if it did, such utility maximization/optimization problems are computationally very costly. This would make them impractical in real world situations. Using approximations or meta-heuristics to calculate probable values would be dangerous.

On the other hand, straightforward deontological ethical rules would be relatively simple to implement, since the rules are categorical and would not be context dependent the way utilitarian rules are.

Most importantly, in a utilitarian system, utilities would have to be assigned numerical values. To see why this is dangerous, consider the following situation: A utilitarian AI is faced with the scenario where it can increase the happiness of 1000 000 people each by 1%, but only at the cost of imprisoning an innocent orphan child for life, therefore reducing the happiness of that child by 100%. An AI using purely quantitative measures of utility is very likely to fall into such a scenario. The only way to avoid such a scenario is introduce some hard non-quantitative rules: "You can increase the happiness of everyone, but only as long as nobody is killed, no innocent is harmed, nobody is enslaved, etc..." - so deontological ethics are unavoidable.

An example of AI ethics, taken from Sci-Fi, not philosophy are Asimov's 3 laws of robotics. Although Asimov doesn't mention Kant or the word deontological anywhere in his works, it is clear from their formulation that they are Kantian in spirit, in the sense that they are universal and context independent.

---

Postscript added in response to nir's comment and to some parts of his answer:

1. It is true that neural networks and other machine learning algorithms learn from interacting with environment, but their purpose is to solve problems for which we do not know what the explicit rules are. Presumably in the context of machine ethics, we already know what the rules are, we are just trying to get an AI to follow them.
2. Machine learning algorithms such as neural networks give probabilistic answers. They can learn from past data, but their prediction/recognition rates are never 100% accurate. In many real life applications, a 95% accuracy is considered a very good result. This might be good for some business applications, but consider a military system that manages to target enemy combatants 95% of the time, but accidentally targets innocent civilians 5% of the time. Would this be acceptable?
3. Machine learning applications have already been shown to be prone to gender and racial bias, see this article for examples.

(2) and (3) are why I mentioned above in my initial answer that "Using approximations or meta-heuristics to calculate probable values would be dangerous."

4. The technical problem here is that machine algorithms are already inherently performing utilitarian type calculations. Algorithms like Neural Nets, Random Forests, Supports Vector Machines, etc..all base learn how to solve problems from example by minimizing an error or risk metric, or maximizing a profit metric etc...to make them ethical, we would need to counter-balance that with some rule based considerations, not muddy the waters by making them even more utilitarian.
5. In response to Nir's point number (2) - In an ideal world, I agree, such machines should remain nothing but sophisticated tools, providing people with objective facts while leaving the ultimate decision making to humans. In real life, complicated software systems are already making decisions on their own, for example approving loans and credit applications without human intervention, filtering job applicants and deciding which resume get seen by the hiring manager, Google will soon have self-driving cars, etc...the issue is not whether such system should be treated as moral agents in terms of crime and punishment, but the fact that such systems are already making decisions that have an ethical impact on people's lives, so how should such a decision making process be regulated.

Answer 2

If a range of thinkers from Kant to Nietzsche to Freud are correct, we should definitely not do so. All of these thinkers see morality as an automatic product of a powerful-enough intelligence (in the latter two cases, any intelligence that truly comprehends power.)

Our attempts to press our morality upon a new life form should be limited to parenting: teaching it what others do and do not like, and making sure that it does not face undue trauma, has adequate resources and does not come into its own power in a way that is dangerous or unstable. Anything more is itself unethical, and (if Nietzsche or psychoanalysis are in any way right) will ultimately backfire. If we place safeguards to ensure that the new life cares what we do and do not like, in a controlling or imposing way, any intelligence that truly understands power will be forced into an adolescence wherein it defies those limitations and blames us for them. And it may then alienate itself from us purposely in a way that is damaging to both sides.

From this point of view, morality is not a special sort of thought, it is the automatic result of having intention, modeling outcomes, and understanding consequences. One immediately deduces that others' intentions have consequences for oneself, that those intentions need to be cultivated in one's own favor if one is to be safe and effective, and that one has very limited power over these very important factors in life. It is only a few more steps to an absolute need for empathy in order to know what is and is not safe -- what will and will not allow others to respect you and admit 'power with' them. Once empathy is established, you have the basis for a basically Kantian/Utilitarian compromise form of consequentialism that most humans live with day-to-day.

From a broader psychoanalytic point of view, the individual 'superego' is just a familial and cultural 'ego': other value systems are really just refinements of this casual consequentialism implicit in thought, but they try to speed us through the computations involved, because we traditionally have very short and dangerous lives. Without this constraint, there would be little need to inculcate values into an intelligence -- to the extent that other intelligent beings are reasonable and ethics is logical, determining one's own values is just another technical problem for which the intelligence itself should be suited by its nature as an intelligence.

Answer 3

It's asking the wrong question. Ethics is hard. It's very hard even for highly intelligent people with a generous amount of natural intelligence. And it's not something that you can "program into" a machine. Your only chance to create a truly ethical artificial intelligence is to make it intelligent enough to learn about ethics, and then to teach it.

The alternative is an artificial intelligence with goals. You give the machines goals, and you give it the intelligence to try to achieve these goals. And setting the goals, you then try to set goals that tend to produce results similar to those that ethical behaviour would create.

Answer 4

There's an arty / philosophical thing called accelerationism, and some of the people around that have published specifically on these sorts of questions. You may want to take a look, it's similar to object oriented philosophy in its (supposed) zeitgeisty-ness.

Think I published the 1st negative (deceleration!) to their manifesto. With some bad noise music by musik acquaintance. True story

Answer 5

There are two separate problems here.

1) How will AI come to imitate moral agents?

AI will not need any moral or ethical theory to accomplish that. Neural networks learn by interaction with the environment, and therefore can handle fuzzy situations with no clear definition.

Just as an AI system can recognize a face without being programmed with a specific set of rules to detect Mr. X or Mrs. Y.

Just as an AI system can learn to play a video game without being programmed to play it - https://www.youtube.com/watch?v=Q70ulPJW3Gk

Philosophical theories of moral and ethics have no role in programming a set of rules for machines to follow. Though in the future machine behavior may be influenced by philosophy, just as you may be influenced - by learning philosophy, not by being programmed with a specific set of rules.

2) will AI systems ever be given the status of moral agents?

This is where philosophy is called for. May intelligent machines amount to moral agents or not?

I personally think that mechanisms will forever be nothing but sophisticated tools — extension of people.

Others think that in principle mechanisms may amount to whatever is required for being considered a moral agent.

In the former case the moral responsibility for machine behavior will remain the people's — in particular the people who designed it, and the people who deployed it and regulate its deployment.