We should take moral advice from computers and not the other way around

>To make a case otherwise is to claim one of two things: either that humans have access to morally relevant data, which is in some way fundamentally inaccessible to computers, or that humans can engage in a kind of moral reasoning which is fundamentally uncomputable.

I remember some quotes about metrics and stats that are something like:

"What is easily counted will be deemed important, what is difficult to count will be deemed unimportant."

Regardless of the intent the data is going to tilt AI in some directions purely based on the data it has and at the exclusion of other information.

And of course Goodhart's law:

"When a measure becomes a target, it ceases to be a good measure."

I'll admit that I don't fully understand the entire paper but it would seem to assume that computers would just magically have all the perfect information, and/or human behavior would not change to skew the results...

Many of us tend to think of morality as a metaphysical issue and believe that a moral sense is something that we as humans have that machines do not.

But this proposal skirts around that, because it does not require the machine to have a moral sense. It merely requires the machine to be able to enforce logical consistency.

So the machine says that IF you as a human want to achieve Ethical Principle A, THEN here are some actions you can take are known to help you achieve that principle.

Crucially, it does not require the machine to perceive Ethical Principle A as true and make a case for why it's true.

So it's very much an augmented morality device, rather than a machine that proves moral truths.

Or to put it another way, it doesn't say, "You should do X because after running a detailed analysis I've determined X is ethical." It says, "You've determined X is ethical, and here are some ways to achieve X."

In the U.S. at least, most people consider it morally outrageous to eat a dog. And yet, many of the same people will eat a hamburger. While there are obvious cultural/emotional reasons people feel this way, I can't think of any logical reason why eating a cow is less troubling than eating a dog.

Put simply, most popular moral frameworks people follow seem full of contradictions like this. And yet, it seems like any attempt to force a different moral framework on people would be a tyrannical force of misery. Any resolution a computer could offer to the contradiction above would make a lot of people deeply unhappy, which is in and of itself a moral problem.

People are emotionally driven, and I'm not sure there exists any logical set of rules that could define a morality people are happy with.

Human interactions are complicated, and ethics is hard. This is why economists have jobs.

Consider the eminent domain problem: a government is considering taking someone's farm to build a road. Relevant facts include: how many people would benefit from using the road, and by how much? (Enormously complex question -- one has to model the rearrangement of businesses and homes as a result of the road's existence.) How much does the citizenry value property rights? (Enormously complex -- how do you quantify the benefit of property rights in the same terms as the benefit from physical goods and services? How do you model the effect of the precedent set by this particular case on future cases?) What is the value of the home in question? (Slightly easier but still not simple: Can you just compare this farm to neighboring farms? What if it's special, because the land is different, or its history, or its brand recognition, or for simple sentimental reasons?)

Outside of certain narrow, prescribed domains, I'm sure artificial ethics would require artificial general intelligence. When I see academic economists being replaced by computers, I'll believe it's within reach.

I don't know if OP is the author, but if so could you suggest some papers where people actually go about implementing ethics into a system even just as a very basic PoC? I'd be really interested. I've looked for papers but even in a survey completed as recently as 2018 it suggested the closest thing we have is a graph database sitting on top of the Moral Machine (which is a survey of ethical questions).

Could we please get that changed to the actual article and add a (2016)?

http://blog.practicalethics.ox.ac.uk/2016/02/oxford-uehiro-p...

That contains much more debatable content. But even with the linked excerpt, I find it hard to draw the rather absolute conclusion the title suggests. Computers are indeed very good at both points, but only after a human carefully crafted a problem specific set of instructions to hold its virtual hand.

I think the bias and modelling discussions of the last few years sufficiently show that it's rather non-trivial to project these functions to algorithms/ML/... and state of the art in the foreseeable future is not particularly good at many aspects of the tasks mentioned in the article.

In the long form it even draws analogies between asking the phone for food recommendations and ethics considerations. Every recommendation engine out there is being actively gamed, that was the case in 2016 just as it is now. There's some truth to the article of course, e.g. more crafted systems that actively address human biases, but we're nowhere near a point to even remotely let our computers take over ethical considerations - even if it's just a recommendation.

So more statistical processing and comparisons in the end, and the final decision is still left to the user? Seems like the proverbial Elves in Tolkien's world where it is said "go not to them for advice as they'll say both yay and nay..."

> After all, a common complaint about practical ethics is that there are too many factors for any one person to consider. A similar complaint could be made about weather forecasting, economic modelling and space travel – and yet we seem to be able to do all of the above just fine with the aid of computers.

Space travel is very hard, but it's "just physics." Weather forecasting is harder, and we get it wrong a lot of the time. Economic modelling is harder still, and our inability to do this accurately was arguably a major contributor to the subprime mortgage issue that led to the last economic down-turn. So, taking moral advice from computers seems like a bad idea.

I think it helps to distinguish between which problems are "wicked" or not - while it may be technically true that some function is represented by a Turing machine, as a practical matter the amount of data would be too large, and how one measures the "output" of the function has so much ambiguity, that it's not practically doable. (I don't know the original source of the term "wicked problem," but it's a concept others have written a lot about why it's different.)

Machine learning does nothing more than reenforce existing statistics, and that is completely amoral.

This very day in the USA, computers are deciding whether prisoners are eligible for parole, and the biggest statistical determiner burdening them is their skin color. That goes completely against the morals of a country holding notions of free, equal citizens and due process.

Anyone considering this question should read Joseph Weizenbaum's 1976 book, Computer Power and Human Reason.

He considers if an AI could be a good judge, and decides that it would not. Even though all the facts, rules and heuristics could be fed into an AI judge, the element we would not be able to program is human compassion. He concludes that even though there are some tasks a computer might be able to do, there will always remain some decisions that should always be made by humans.

i think that due to the fact all humans carry imperfect information, their opinions on good/bad/ethical varies, and that will ensure that never a system can be made which 'behaves 100% ethically'. it might behave in such a way with respect to the opinion of the person claiming that, but to another person it can seem completely unreasonable/unethical...

since you can't create an AI which takes into account all the flaws of all humans present in their knowledge and consciousness, a system which has such perfection is impossible to make. (even these flaws are often just perceived flaws and them being a flaw is an subjective matter based upon other subjective matters.)

it's not about having all the data for an AI, it's more about understanding what lack of data means to humans and how it affects their decision and interpretations, and about how the same data can be interpreted in many ways.

Even if all humans were exposed to the same data as each other exactly, they would likely still carry different opinions and interpret the same data differently leading to a completely different decision making process... I think this at the moment is inherently impossible to create within or take into account in current computers or programming.

If you would reverse it, and have humans take all their morals and ethics from computers, what is left of humans? Isn't that what makes a human? the ability and/or inability to do this themselves. i think no one is looking for a world or working towards a world where only 1 human exists in multitude. i think the work should be focused on preserving the uniqueness of identity while maximising its potential within that uniqueness. That also makes me of the opinion that AI should thus be specialised within domains of operation, and not attempted to be implemented in a general fasion.

perhaps an AI system could exist which comprises of many specific AI systems, which would make it more generally applicable based upon many input from specialised AI systems, who knows. But 1 system and 1 data set will never be able to cover inherent uniqueness within humans.

you can argue about some rotten apple humans who have 'bad behaviour', but even the good people you know, are wildely different from you. admit it. you are not them, and they are not you and that's how it should be.

The link here goes to a page that provides just the very beginning of the essay and a bit of context, along with a link to http://blog.practicalethics.ox.ac.uk/2016/02/oxford-uehiro-p... where you can find the rest of it.

Computer: there is a small group of unpopular people causing social problems. Liquidating them humanely would increase net happiness at the least cost.

Shall we proceed?

I read this as "We should take moral advice from the people who program the computers and not the other way around"

Very interesting read, but I strongly disagree with it. I could mention several technical issues, but I do not think it would lead do a fruitful discussion.

Instead, think about this: why would people not behaving ethically? The point of the essay seems to be that they are either not aware of what is the best thing to do, or what the consequences of certain actions would be.

I would argue that no, (most) people do know, but for some reason would still behave unethically. Who needs a reminder that driving while drunk puts lives in danger? Nobody. It is well known that more elaborate/advanced ethical reasoning does not lead to more ethical behavior.

And most ethical dilemmas are not that simple, consider:

> If the output of the computer ever contradicts with the human’s own intuitions, the human would be provided with an example of how and why the moral framework they have chosen does not match up with what they consider to be moral.

A human could as well come up with several arguments for why their intuition is better than the computer's advice. Because that's how ethics is, there is not a single right answer, there are many good possibilities and many bad ones. Humans do not need help in identifying the bad ones, and only an oracle can say which alternative is the "best".

In fact, this seems to be the main purpose of such an ethical assistant, to "remove important epistemic limitations", i.e. predict the future.

> Suppose every driver had access to a good estimate for how likely they were to be involved in a traffic accident, every time they decided to go out.

This seems utterly impossible. Most accidents are caused by random events, and the only reasonable estimate for this probability is between zero and a number so tiny that it would not make any practical difference.

TLDR:

- I am not going to argue whether computers might be able to perform decent moral reasoning

- But if they can, it cannot be much better than what humans can already do

- And even if it was better, humans would likely not follow this advice

- Simply because people sometimes do not want to behave rationally