Followup toThe Bedrock of Fairness

Every time I wonder if I really need to do so much prep work to explain an idea, I manage to forget some minor thing and a dozen people promptly post objections.

In this case, I seem to have forgotten to cover the topic of how morality applies to more than one person at a time.

Stop laughing, it's not quite as dumb an oversight as it sounds.  Sort of like how some people argue that macroeconomics should be constructed from microeconomics, I tend to see interpersonal morality as constructed from personal morality.  (And definitely not the other way around!)

In "The Bedrock of Fairness" I offered a situation where three people discover a pie, and one of them insists that they want half.  This is actually toned down from an older dialogue where five people discover a pie, and one of them—regardless of any argument offered—insists that they want the whole pie.

Let's consider the latter situation:  Dennis wants the whole pie.  Not only that, Dennis says that it is "fair" for him to get the whole pie, and that the "right" way to resolve this group disagreement is for him to get the whole pie; and he goes on saying this no matter what arguments are offered him.

This group is not going to agree, no matter what.  But I would, nonetheless, say that the right thing to do, the fair thing to do, is to give Dennis one-fifth of the pie—the other four combining to hold him off by force, if necessary, if he tries to take more.

A terminological note:

In this series of posts I have been using "morality" to mean something more like "the sum of all values and valuation rules", not just "values that apply to interactions between people".

The ordinary usage would have that jumping on a trampoline is not "morality", it is just some selfish fun.  On the other hand, giving someone else a turn to jump on the trampoline, is more akin to "morality" in common usage; and if you say "Everyone should take turns!" that's definitely "morality".

But the thing-I-want-to-talk-about includes the Fun Theory of a single person jumping on a trampoline.

Think of what a disaster it would be if all fun were removed from human civilization!  So I consider it quite right to jump on a trampoline.  Even if one would not say, in ordinary conversation, "I am jumping on that trampoline because I have a moral obligation to do so."  (Indeed, that sounds rather dull, and not at all fun, which is another important element of my "morality".)

Alas, I do get the impression that in a standard academic discussion, one would use the term "morality" to refer to the sum-of-all-valu(ation rul)es that I am talking about.  If there's a standard alternative term in moral philosophy then do please let me know.

If there's a better term than "morality" for the sum of all values and valuation rules, then this would free up "morality" for interpersonal values, which is closer to the common usage.

Some years ago, I was pondering what to say to the old cynical argument:  If two monkeys want the same banana, in the end one will have it, and the other will cry morality.  I think the particular context was about whether the word "rights", as in the context of "individual rights", meant anything.  It had just been vehemently asserted (on the Extropians mailing list, I think) that this concept was meaningless and ought to be tossed out the window.

Suppose there are two people, a Mugger and a Muggee.  The Mugger wants to take the Muggee's wallet.  The Muggee doesn't want to give it to him.  A cynic might say:  "There is nothing more to say than this; they disagree.  What use is it for the Muggee to claim that he has an individual_right to keep his wallet?  The Mugger will just claim that he has an individual_right to take the wallet."

Now today I might introduce the notion of a 1-place versus 2-place function, and reply to the cynic, "Either they do not mean the same thing by individual_right, or at least one of them is very mistaken about what their common morality implies."  At most one of these people is controlled by a good approximation of what I name when I say "morality", and the other one is definitely not.

But the cynic might just say again, "So what?  That's what you say.  The Mugger could just say the opposite.  What meaning is there in such claims?  What difference does it make?"

So I came up with this reply:  "Suppose that I happen along this mugging.  I will decide to side with the Muggee, not the Mugger, because I have the notion that the Mugger is interfering with the Muggee's individual_right to keep his wallet, rather than the Muggee interfering with the Mugger's individual_right to take it.  And if a fourth person comes along, and must decide whether to allow my intervention, or alternatively stop me from treating on the Mugger's individual_right to take the wallet, then they are likely to side with the idea that I can intervene against the Mugger, in support of the Muggee."

Now this does not work as a metaethics; it does not work to define the word should.  If you fell backward in time, to an era when no one on Earth thought that slavery was wrong, you should still help slaves escape their owners.  Indeed, the era when such an act was done in heroic defiance of society and the law, was not so very long ago.

But to defend the notion of individual_rights against the charge of meaninglessness, the notion of third-party interventions and fourth-party allowances of those interventions, seems to me to coherently cash out what is asserted when we assert that an individual_right exists.  To assert that someone has a right to keep their wallet, is to assert that third parties should help them keep it, and that fourth parties should applaud those who thus help.

This perspective does make a good deal of what is said about individual_rights into nonsense.  "Everyone has a right to be free from starvation!"  Um, who are you talking to?  Nature?  Perhaps you mean, "If you're starving, and someone else has a hamburger, I'll help you take it."  If so, you should say so clearly.  (See also The Death of Common Sense.)

So that is a notion of individual_rights, but what does it have to do with the more general question of interpersonal morality?

The notion is that you can construct interpersonal morality out of individual morality.  Just as, in this particular example, I constructed the notion of what is asserted by talking about an individual_right, by making it an assertion about whether third parties should decide, for themselves, to intefere; and whether fourth parties should, individually, decide to applaud the interference.

Why go to such lengths to define things in individual terms?  Some people might say:  "To assert the existence of a right, is to say what society should do."

But societies don't always agree on things.  And then you, as an individual, will have to decide what's right for you to do, in that case.

"But individuals don't always agree within themselves, either," you say.  "They have emotional conflicts."

Well... you could say that and it would sound wise.  But generally speaking, neurologically intact humans will end up doing some particular thing. As opposed to flopping around on the floor as their limbs twitch in different directions under the temporary control of different personalities.  Contrast to a government or a corporation.

A human brain is a coherently adapted system whose parts have been together optimized for a common criterion of fitness (more or less).  A group is not functionally optimized as a group.  (You can verify this very quickly by looking at the sex ratios in a maternity hospital.)  Individuals may be optimized to do well out of their collective interaction—but that is quite a different selection pressure, the adaptations for which do not always produce group agreement!  So if you want to look at a coherent decision system, it really is a good idea to look at one human, rather than a bureaucracy.

I myself am one person—admittedly with a long trail of human history behind me that makes me what I am, maybe more than any thoughts I ever thought myself.  But still, at the end of the day, I am writing this blog post; it is not the negotiated output of a consortium.  It is quite easy for me to imagine being faced, as an individual, with a case where the local group does not agree within itself—and in such a case I must decide, as an individual, what is right.  In general I must decide what is right!  If I go along with the group that does not absolve me of responsibility.  If there are any countries that think differently, they can write their own blog posts.

This perspective, which does not exhibit undefined behavior in the event of a group disagreement, is one reason why I tend to treat interpersonal morality as a special case of individual morality, and not the other way around.

Now, with that said, interpersonal morality is a highly distinguishable special case of morality.

As humans, we don't just hunt in groups, we argue in groups.  We've probably been arguing linguistically in adaptive political contexts for long enough—hundreds of thousands of years, maybe millions—to have adapted specifically to that selection pressure.

So it shouldn't be all that surprising if we have moral intuitions, like fairness, that apply specifically to the morality of groups.

One of these intuitions seems to be universalizability.

If Dennis just strides around saying, "I want the whole pie!  Give me the whole pie!  What's fair is for me to get the whole pie!  Not you, me!" then that's not going to persuade anyone else in the tribe.  Dennis has not managed to frame his desires in a form which enable them to leap from one mind to another.  His desires will not take wings and become interpersonal.  He is not likely to leave many offspring.

Now, the evolution of interpersonal moral intuitions, is a topic which (he said, smiling grimly) deserves its own blog post.  And its own academic subfield.  (Anything out there besides The Evolutionary Origins of Morality?  It seemed to me very basic.)

But I do think it worth noting that, rather than trying to manipulate 2-person and 3-person and 7-person interactions, some of our moral instincts seem to have made the leap to N-person interactions.  We just think about general moral arguments.  As though the values that leap from mind to mind, take on a life of their own and become something that you can reason about.  To the extent that everyone in your environment does share some values, this will work as adaptive cognition.  This creates moral intuitions that are not just interpersonal but transpersonal.

Transpersonal moral intuitions are not necessarily false-to-fact, so long as you don't expect your arguments cast in "universal" terms to sway a rock.  There really is such a thing as the psychological unity of humankind.  Read a morality tale from an entirely different culture; I bet you can figure out what it's trying to argue for, even if you don't agree with it.

The problem arises when you try to apply the universalizability instinct to say, "If this argument could not persuade an UnFriendly AI that tries to maximize the number of paperclips in the universe, then it must not be a good argument."

There are No Universally Compelling Arguments, so if you try to apply the universalizability instinct universally, you end up with no morality.  Not even universalizability; the paperclip maximizer has no intuition of universalizability.  It just chooses that action which leads to a future containing the maximum number of paperclips.

There are some things you just can't have a moral conversation with.  There is not that within them that could respond to your arguments.  You should think twice and maybe three times before ever saying this about one of your fellow humans—but a paperclip maximizer is another matter.  You'll just have to override your moral instinct to regard anything labeled a "mind" as a little floating ghost-in-the-machine, with a hidden core of perfect emptiness, which could surely be persuaded to reject its mistaken source code if you just came up with the right argument.  If you're going to preserve universalizability as an intuition, you can try extending it to all humans; but you can't extend it to rocks or chatbots, nor even powerful optimization processes like evolutions or paperclip maximizers.

The question of how much in-principle agreement would exist among human beings about the transpersonal portion of their values, given perfect knowledge of the facts and perhaps a much wider search of the argument space, is not a matter on which we can get much evidence by observing the prevalence of moral agreement and disagreement in today's world.  Any disagreement might be something that the truth could destroydependent on a different view of how the world is, or maybe just dependent on having not yet heard the right argument.  It is also possible that knowing more could dispel illusions of moral agreement, not just produce new accords.

But does that question really make much difference in day-to-day moral reasoning, if you're not trying to build a Friendly AI?

 

Part of The Metaethics Sequence

Next post: "Morality as Fixed Computation"

Previous post: "The Meaning of Right"

New Comment
30 comments, sorted by Click to highlight new comments since:

Re: Anything out there besides The Evolutionary Origins of Morality?

Well you read "The Origins of Virtue" IIRC. That has 18 pages of references and notes.

...and how about "The Moral Animal"?

Mr. Yudkowsky says: But does that question really make much difference in day-to-day moral reasoning, if you're not trying to build a Friendly AI?

Here is a quote that I think speaks to that:

The Singularity is the technological creation of smarter-than-human intelligence. Ask what “smarter-than-human” really means. And as the basic definition of the Singularity points out, this is exactly the point at which our ability to extrapolate breaks down. We don’t know because we’re not that smart.-- Eliezer Yudkowsky

As I understand it, it is not possible for a human to design a machine that is "smarter-than-human", by definition. It is possible, however to design a machine that can design a machine that is smarter than human. According to one of my correspondents, it has already occurred (quoted exactly, grammar and all):

"My current opinion is that the singularity is behind us. The deep discovery is the discovery of the Universal Machine, alias the computer, but we have our nose so close on it that we don't really realize what is happening. From this, by adding more and more competence to the universal machine, we put it away from his initial "natural" intelligence. I even believe that the greek theologians were in advance, conceptually, on what intelligence his. Intelligence is confused with competence today. It is correct that competence needs intelligence to develop, but competence cannot be universal and it makes the intelligence fading away: it has a negative feedback on intelligence. So my opinion is that the singularity has already occurred, and, since a longer time, we have abandon the conceptual tools to really appreciate that recent revolution. We are somehow already less smarter than the universal machine, when it is not yet programmed."

Best regards,

Bruno Marchal IRIDIA-ULB http://iridia.ulb.ac.be/~marchal/

[-]PK240

What is "morality" for?(not morality) The "morality" concept seems so slippery at this point that it might be better to use other words to more clearly communicate meaning.

In any case, be it personal or interpersonal morality, it is more efficient to agree on a single goal and then optimize towards it. The power of optimization is in driving the environment in the same direction, multiplying the effect with each interaction. For personal morality, it means figuring out morality more accurately than you needed before, perhaps more accurately than is possible from mere intuition and argument, and in particular balancing the influences of conflicting drives. For interpersonal morality, it means agreeing on global target, even if it's suboptimal for each party considered separately.

closing the tag...

[-]PK200

To clarify my question, what is the point of all this talk about "morality" if it all amounts to "just do what you think is right"? I mean other than the futility of looking for The True Source of morality outside yourself. I guess I may have answered my own question if this was the whole point. So now what? How do I know what is moral and what isn't? I mean I can answer the easy question but how do I solve the hard ones? I was expecting to get easy answers to moral questions from your theory Eliezer. I feel cheated now.

I mean other than the futility of looking for The True Source of morality outside yourself

The argument that morality must have a Source is conditioned by the way most people encounter moral exhortation, thtough religion. The idea is that moral principles are compulsions or obligations, which means, in turn, that they are something like edicts or commandments. Whilst there are lots of humans who can offer opinions on what is compulsory, they are all subjective opinions: no-one can lay down an objective edict. However, the basic idea of a "source" merges two separate ideas: who tells you about some item of knowledge, and what makes it true. The mathematical truths relayed to you by your teachers are objectively true, not because teachers are Objective Sources, but because of the ways they are proved an justified — which is not by the teacher personally. The set of things tha are true just because someone says so is small. Eye witness testimony, personal reminiscences, reports of subjective states: not many things work on the Source system, and there is no reason to assume ethics does.

PK: Not "do what you think is right", but "deliberate the way you think is right" - keep aiming for the ideal implicit in and relative to, but not currently realized in, yourself. Definitely not a free lunch, no.

As I understand it, it is not possible for a human to design a machine that is "smarter-than-human", by definition.

Then we couldn't design a machine that could design a machine that would be smarter than we were, either. Machine-2's design couldn't be smarter than it, and machine-2 couldn't be smarter than machine-1 which designed it, which in turn couldn't be smarter than the designers of machine-1: us.

We can't hold in our individual minds a design that is more complex than one of those minds, or in our collective minds what is more complex than the collective. But the design doesn't have to be as complex as the thing that is designed, and the representation of the design is simpler still.

It's trivially easy for a human being to design a data-encoding-and-storage system that can hold more data than is contained in that human's brain. The brain just can't represent that system.

Re: As I understand it, it is not possible for a human to design a machine that is "smarter-than-human", by definition.

Maybe not - but a bunch of humans could probably do it.

Re: My current opinion is that the singularity is behind us.

See my essay: The Intelligence Explosion Is Happening Now.

sophiesdad: As I understand it, it is not possible for a human to design a machine that is "smarter-than-human", by definition.

Caledonian: Your understanding is mistaken.

Mr. Caledonian, I'm going to stick by my original statement. Smarter-than-human intelligence will be designed by machines with "around human" intelligence running recursive self-improvement code. It will not start with a human-designed superhuman intelligence. How could a human know what that is? That's why I'm not sure that all the years of thought going into the meaning of morality, etc. is helpful. If it is impossible for humans to understand what superhuman intelligence and the world around it would be like, just relax and go along for the ride. If we're destroyed, we'll never know it. If you're religious, you'll be in heaven (great) or hell (sucks). I agree with Tim Tyler (who must not be drinking as much as when he was commenting on Taubes) that we already have machines that perform many tasks, including design tasks, that would be impractical for humans alone.

Smarter-than-human intelligence will be designed by machines with "around human" intelligence running recursive self-improvement code. It will not start with a human-designed superhuman intelligence. How could a human know what that is?

How could the machine know what superhuman intelligence is? If the machine can design machines that are smarter than it, precisely why can't humans design machines smarter than them?

I don't think Deep Blue "knew" that it was trying to beat Gary Kasparov in the game of chess. It was programmed to come up with every possible alternative move and evaluate the outcome of each in terms of an eventual result of taking K's king. The human brain is elegant, but it's not fast, and unquestionable no human could have evaluated all the possible moves within the time limit. Deep Blue is quaint compared to the Universal Machines of the near future. David Deutsch claims that quantum computers will be able to factor numbers that would require a human more time than is known to have existed in the history of the universe. It won't have superhuman intelligence, but it will be fast. Imagine if it's programs were recursively self-improving.

If there's a standard alternative term in moral philosophy then do please let me know.

As far as I know, there is not. In moral philosophy, when deontologists talk about morality, they are typically talking about things that are for the benefit of others. Indeed, they even have conversations about how to balance between self-interest and the demands of morality. In contrast, consequentialists have a theory that already accounts for the benefit of the agent who is doing the decision making: it counts just as much as anyone else. Thus for consequentialists, there is typically no separate conflict between self-interest and morality: morality for them already takes this into account. So in summary, many moral philosophers are aware of the distinction, but I don't know of any pre-existing terms for it.

By the way, don't worry too much about explaining all pre-requisites before making a post. Explaining some of them afterwards in response to comments can be a more engaging way to do it. In particular, it means that us readers can see which parts we are skeptical of and then just focus our attention on posts which defend that aspect, skimming the ones that we already agree with. Even when it comes to the book, it will probably be worth giving a sketch of where you want to end up early on, with forward references to the appropriate later chapters as needed. This will let the readers read the pre-requisite chapters in a more focused way.

As far as I know, there is not. In moral philosophy, when deontologists talk about morality, they are typically talking about things that are for the benefit of others.

Typically, maybe, but not necessarily. There is no obvious contradiction in the idea of a rule of self-preservation or self-enhancement. Many consider suicide imorroral, for instance.

Indeed, they even have conversations about how to balance between self-interest and the demands of morality. In contrast, consequentialists have a theory that already accounts for the benefit of the agent who is doing the decision making: it counts just as much as anyone else.

ie one 6 billionth in the case of humans.

I imagine you're probably aware of this in the meantime, but for Eliezer's benefit in case he isn't (and hopefully for the benefit of others who read this post and aren't as familiar with moral philosophy): I believe the term "normativity" is the standard term used to refer to the "sum of all valu(ation rul)es," and would probably be a good term for LessWrong to adopt for this purpose.

I would expect that libertarians' utility comes unexpectedly close to what Mr. Yudkowsky calls morality.

I would find it very difficult to summarize what I had not written. Any attempt at summary would turn into the whole post. It's how my authorness works. Part of the whole point of writing this on Overcoming Bias is so that I don't have to write it all into the book, and can just put a footnote somewhere with a clean conscience.

I found this post more engaging than the last, and the first half genuinely instructive. I had a vague idea about what social morality would have to look like, but this is actually a working theory. IMHO, it's still a kind of depressing outlook--there's a reason so many find the quest for an objective morality so appealing--but much better than nihilism. And I'm easily depressed. :)

Re: why can't humans design machines smarter than them

The claim was: it is not possible for a human to design a machine that is "smarter-than-human".

Billions of humans collectively haven't managed to construct a superintelligence over the last 50 years. It seems unlikely that a single human could do it - the problem is too big and difficult.

But I would, nonetheless, say that the right thing to do, the fair thing to do, is to give Dennis one-fifth of the pie - the other four combining to hold him off by force, if necessary, if he tries to take more.
What do the words 'right' and 'fair' mean in this sentence?

Don't 'taboo' the words. 'Tabooing' only requires that the specific word not be used. The rules of 'tabooing' don't require that you convey meaning; they let you get away with slapping a new label over the old one without ever taking a look at what's underneath it.

WHAT DO THOSE WORDS MEAN?

If you are confused by a problem and don't know how to solve it, you do not know how much remaining effort it will take to solve once you are unconfused, or what it will take to unconfuse yourself. You might as easily point out that "billions of humans have failed to do X over the least 50 years" for all X such that it hasn't happened yet.

Imagine the year 2100

AI Prac Class Task: (a) design and implement a smarter-than-human AI using only open source components; (b) ask it to write up your prac report. Time allotted: 4 hours Bonus points: disconnect your AI host from all communications devices; place your host in a Faraday cage; disable your AI's morality module; find a way to shut down the AI without resorting to triggering the failsafe host self-destruct.

sophiesdad, since a human today could not design a modern microprocessor (without using the already-developed plethora of design tools) then your assertion that a human will never design a smarter-than-human machine is safe but uninformative. Humans use smart tools to make smarter tools. It's only reasonable to predict that smarter-than-human machines will only be made by a collaboration of humans and existing smart machines.

Speculation on whether "smart enough to self improve" comes before or after the smart-as-a-human mark on some undefined 1-dimensional smartness scale is fruitless. By the look of what you seem to endorse by quoting your unnamed correspondent, your definition of "smart" makes comparison with human intelligence impossible.

disable your AI's morality module; find a way to shut down the AI without resorting to triggering the failsafe host self-destruct.

Trivial. Once you've disabled your AI's morality module, you've already shut it down.

You just build the conscience, and that is the AI.

[-]ata140

But to defend the notion of individual_rights against the charge of meaninglessness, the notion of third-party interventions and fourth-party allowances of those interventions, seems to me to coherently cash out what is asserted when we assert that an individual_right exists. To assert that someone has a right to keep their wallet, is to assert that third parties should help them keep it, and that fourth parties should applaud those who thus help.

This perspective does make a good deal of what is said about individual_rights into nonsense. "Everyone has a right to be free from starvation!" Um, who are you talking to? Nature?

Why such an uncharitable interpretation, particularly when the more plausible unpacking is exactly the one you used in the previous sentence? "To assert that someone has a right to be free from starvation, is to assert that third parties should help them avoid starvation, and that fourth parties should applaud those who thus help." That doesn't sound nonsensical or even incorrect. It certainly doesn't require that you imagine yourself convincing Nature to stop starving people, just as you can say "this potential mugging victim has the right to keep their wallet" without requiring that it be possible to convince the mugger that they shouldn't take the muggee's wallet; even if the mugger is a sociopath or a robot or an alien who couldn't even conceivably be convinced by a moral argument, that would seem to be irrelevant to any argument that third parties should help the victim keep their wallet, and that fourth parties should applaud those who help.

There really is such a thing as the psychological unity of humankind. Read a morality tale from an entirely different culture; I bet you can figure out what it's trying to argue for, even if you don't agree with it.

I'm not sure how the first sentence there relates to the second. For example, we can figure out in great detail what a highly alien "intelligence" like evolution "wants", but that doesn't mean that there's much in common between the evolutionary process and our own mental processes.

[-][anonymous]40

I'm not sure if I fully groked this post. After some thought I came up with this:

Interpersonal morality is the application of generalized intuitions shared with a specific subset of minds on conduct between persons that can facilitate negotiation and cooperation with them. Which is not only a often useful tool, but something we, because of how we are built, sometimes (often?) want to follow.

Can someone please tell me if this is about right?

Sounds pretty good to me.