This post is about an idea I had around 1997, when I was 16 years old and a freshman computer-science major at Cornell. Back then, I was extremely impressed by a research project called CLEVER, which one of my professors, Jon Kleinberg, had led while working at IBM Almaden. The idea was to use the link structure of the web itself to rank which web pages were most important, and therefore which ones should be returned first in a search query. Specifically, Kleinberg defined “hubs” as pages that linked to lots of “authorities,” and “authorities” as pages that were linked to by lots of “hubs.” At first glance, this definition seems hopelessly circular, but Kleinberg observed that one can break the circularity by just treating the World Wide Web as a giant directed graph, and doing some linear algebra on its adjacency matrix. Equivalently, you can imagine an iterative process where each web page starts out with the same hub/authority “starting credits,” but then in each round, the pages distribute their credits among their neighbors, so that the most popular pages get more credits, which they can then, in turn, distribute to their neighbors by linking to them.
I was also impressed by a similar research project called PageRank, which was proposed later by two guys at Stanford named Sergey Brin and Larry Page. Brin and Page dispensed with Kleinberg’s bipartite hubs-and-authorities structure in favor of a more uniform structure, and made some other changes, but otherwise their idea was very similar. At the time, of course, I didn’t know that CLEVER was going to languish at IBM, while PageRank (renamed Google) was going to expand to roughly the size of the entire world’s economy.
In any case, the question I asked myself about CLEVER/PageRank was not the one that, maybe in retrospect, I should have asked: namely, “how can I leverage the fact that I know the importance of this idea before most people do, in order to make millions of dollars?”
Instead I asked myself: “what other ‘vicious circles’ in science and philosophy could one unravel using the same linear-algebra trick that CLEVER and PageRank exploit?” After all, CLEVER and PageRank were both founded on what looked like a hopelessly circular intuition: “a web page is important if other important web pages link to it.” Yet they both managed to use math to defeat the circularity. All you had to do was find an “importance equilibrium,” in which your assignment of “importance” to each web page was stable under a certain linear map. And such an equilibrium could be shown to exist—indeed, to exist uniquely.
Searching for other circular notions to elucidate using linear algebra, I hit on morality. Philosophers from Socrates on, I was vaguely aware, had struggled to define what makes a person “moral” or “virtuous,” without tacitly presupposing the answer. Well, it seemed to me that, as a first attempt, one could do a lot worse than the following:
A moral person is someone who cooperates with other moral people, and who refuses to cooperate with immoral people.
Obviously one can quibble with this definition on numerous grounds: for example, what exactly does it mean to “cooperate,” and which other people are relevant here? If you don’t donate money to starving children in Africa, have you implicitly “refused to cooperate” with them? What’s the relative importance of cooperating with good people and withholding cooperation with bad people, of kindness and justice? Is there a duty not to cooperate with bad people, or merely the lack of a duty to cooperate with them? Should we consider intent, or only outcomes? Surely we shouldn’t hold someone accountable for sheltering a burglar, if they didn’t know about the burgling? Also, should we compute your “total morality” by simply summing over your interactions with everyone else in your community? If so, then can a career’s worth of lifesaving surgeries numerically overwhelm the badness of murdering a single child?
For now, I want you to set all of these important questions aside, and just focus on the fact that the definition doesn’t even seem to work on its own terms, because of circularity. How can we possibly know which people are moral (and hence worthy of our cooperation), and which ones immoral (and hence unworthy), without presupposing the very thing that we seek to define?
Ah, I thought—this is precisely where linear algebra can come to the rescue! Just like in CLEVER or PageRank, we can begin by giving everyone in the community an equal number of “morality starting credits.” Then we can apply an iterative update rule, where each person A can gain morality credits by cooperating with each other person B, and A gains more credits the more credits B has already. We apply the rule over and over, until the number of morality credits per person converges to an equilibrium. (Or, of course, we can shortcut the process by simply finding the principal eigenvector of the “cooperation matrix,” using whatever algorithm we like.) We then have our objective measure of morality for each individual, solving a 2400-year-old open problem in philosophy.
The next step, I figured, would be to hack together some code that computed this “eigenmorality” metric, and then see what happened when I ran the code to measure the morality of each participant in a simulated society. What would happen? Would the results conform to my pre-theoretic intuitions about what sort of behavior was moral and what wasn’t? If not, then would watching the simulation give me new ideas about how to improve the morality metric? Or would it be my intuitions themselves that would change?
Unfortunately, I never got around to the “coding it up” part—there’s a reason why I became a theorist! The eigenmorality idea went onto my back burner, where it stayed for the next 16 years: 16 years in which our world descended ever further into darkness, lacking a principled way to quantify morality. But finally, this year, just two separate things have happened on the eigenmorality front, and that’s why I’m blogging about it now.
Eigenjesus and Eigenmoses
The first thing that’s happened is that Tyler Singer-Clark, my superb former undergraduate advisee, did code up eigenmorality metrics and test them out on a simulated society, for his MIT senior thesis project. You can read Tyler’s 12-page report here—it’s a fun, enjoyable, thought-provoking first research paper, one that I wholeheartedly recommend. Or, if you’d like to experiment yourself with the Python code, you can download it here from github. (Of course, all opinions expressed in this post are mine alone, not necessarily Tyler’s.)
Briefly, Tyler examined what eigenmorality has to say in the setting of an Iterated Prisoner’s Dilemma (IPD) tournament. The Iterated Prisoner’s Dilemma is the famous game in which two players meet repeatedly, and in each turn can either “Cooperate” or “Defect.” The absolute best thing, from your perspective, is if you defect while your partner cooperates. But you’re also pretty happy if you both cooperate. You’re less happy if you both defect, while the worst (from your standpoint) is if you cooperate while your partner defects. At each turn, when contemplating what to do, you have the entire previous history of your interaction with this partner available to you. And thus, for example, you can decide to “punish” your partner for past defections, “reward” her for past cooperations, or “try to take advantage” by unilaterally defecting and seeing what happens. At each turn, the game has some small constant probability of ending—so you know approximately how many times you’ll meet this partner in the future, but you don’t know exactly when the last turn will be. Your score, in the game, is then the sum-total of your score over all turns and all partners (where each player meets each other player once).
In the late 1970s, as recounted in his classic work The Evolution of Cooperation, Robert Axelrod invited people all over the world to submit computer programs for playing this game, which were then pit against each other in the world’s first serious IPD tournament. And, in a tale that’s been retold in hundreds of popular books, while many people submitted complicated programs that used machine learning, etc. to try to suss out their opponents, the program that won—hands-down, repeatedly—was TIT_FOR_TAT, a few lines of code submitted by the psychologist Anatol Rapaport to implement an ancient moral maxim. TIT_FOR_TAT starts out by cooperating; thereafter, it simply does whatever its opponent did in the last move, swiftly rewarding every cooperation and punishing every defection, and ignoring the entire previous history. In the decades since Axelrod, running Iterated Prisoners’ Dilemma tournaments has become a minor industry, with countless variations explored (for example, “evolutionary” versions, and versions allowing side-communication between the players), countless new strategies invented, and countless papers published. To make a long story short, TIT_FOR_TAT continues to do quite well across a wide range of environments, but depending on the mix of players present, other strategies can sometimes beat TIT_FOR_TAT. (As one example, if there’s a sizable minority of colluding players, who recognize each other by cooperating and defecting in a prearranged sequence, then those players can destroy TIT_FOR_TAT and other “simple” strategies, by cooperating with one another while defecting against everyone else.)
Anyway, Tyler sets up and runs a fairly standard IPD tournament, with a mix of strategies that includes TIT_FOR_TAT, TIT_FOR_TWO_TATS, other TIT_FOR_TAT variations, PAVLOV, FRIEDMAN, EATHERLY, CHAMPION (see the paper for details), and degenerate strategies like always defecting, always cooperating, and playing randomly. However, Tyler then asks an unusual question about the IPD tournament: namely, purely on the basis of the cooperate/defect sequences, which players should we judge to have acted morally toward their partners?
It might be objected that the players didn’t “know” they were going to be graded on morality: as far as they knew, they were just trying to maximize their individual utilities. The trouble with that objection is that the players didn’t “know” they were trying to maximize their utilities either! The players are bots, which do whatever their code tells them to do. So in some sense, utility—no less than morality—is “merely an interpretation” that we impose on the raw cooperate/defect sequences! There’s nothing to stop us from imposing some other interpretation (say, one that explicitly tries to measure morality) and seeing what happens.
In an attempt to measure the players’ morality, Tyler uses the eigenmorality idea from before. The extent to which player A “cooperates” with player B is simply measured by the percentage of times A cooperates. (One acknowledged limitation of this work is that, when two players both defect, there’s no attempt to take into account “who started it,” and to judge the aggressor more harshly than the retaliator—or to incorporate time in any other way.) This then gives us a “cooperation matrix,” whose (i,j) entry records the total amount of niceness that player i displayed to player j. Diagonalizing that matrix, and taking its largest eigenvector, then gives us our morality scores.
Now, there’s a very interesting ambiguity in what I said above. Namely, should we define the “niceness scores” to lie in [0,1] (so that the lowest, meanest possible score is 0), or in [-1,1] (so that it’s possible to have negative niceness)? This might sound like a triviality, but in our setting, it’s precisely the mathematical reflection of one of the philosophical conundrums I mentioned earlier. The conundrum can be stated as follows: is your morality a monotone function of your niceness? We all agree, presumably, that it’s better to be nice to Gandhi than to be nice to Hitler. But do you have a positive obligation to be not-nice to Hitler: to make him suffer because he made others suffer? Or, OK, how about not Hitler, but someone who’s somewhat bad? Consider, for example, a woman who falls in love with, and marries, an unrepentant armed robber (with full knowledge of who he is, and with other options available to her). Is the woman morally praiseworthy for loving her husband despite his bad behavior? Or is she blameworthy because, by rewarding his behavior with her love, she helps to enable it?
To capture two possible extremes of opinion about such questions, Tyler and I defined two different morality metrics, which we called … wait for it … eigenmoses and eigenjesus. Eigenmoses has the niceness scores in [-1,1], which means that you’re actively rewarded for punishing evildoers: that is, for defecting against those who defect against many moral players. Eigenjesus, by contrast, has the niceness scores in [0,1], which means that you always do at least as well by “turning the other cheek” and cooperating. (Though note that, even with eigenjesus, you get more morality credits by cooperating with moral players than by cooperating with immoral ones.)
This is probably a good place to mention a second limitation of Tyler’s current study. Namely, with the current system, there’s no direct way for a player to find out how its partner has been behaving toward third parties. The only information that A gets about the goodness or evilness of player B, comes from A and B’s direct interaction. Ideally, one would like to design bots that take into account, not only the other bots’ behavior toward them, but the other bots’ behavior toward each other. So for example, even if someone is unfailingly nice to you, if that person is an asshole to everyone else, then the eigenmoses moral code would demand that you return the person’s cooperation with icy defection. Conversely, even if Gandhi is mean and hateful to you, you would still be morally obliged (interestingly, on both the eigenmoses and eigenjesus codes) to be nice to him, because of the amount of good he does for everyone else.
Anyway, you can read Tyler’s paper if you want to see the results of computing the eigenmoses and eigenjesus scores for a diverse population of bots. Briefly, the results accord pretty well with intuition. When we look at eigenjesus scores, the all-cooperate bot comes out on top and the all-defect bot on the bottom (as is mathematically necessary), with TIT_FOR_TAT somewhere in the middle, and generous versions of TIT_FOR_TAT higher up. When we look at eigenmoses, by contrast, TIT_FOR_TWO_TATS comes out on top, with TIT_FOR_TAT in sixth place, and the all-cooperate bot scoring below the median. Interestingly, once again, the all-defect bot gets the lowest score (though in this case, it wasn’t mathematically necessary).
Even though the measures acquit themselves well in this particular tournament, it’s admittedly easy to construct scenarios where the prescriptions of eigenjesus and eigenmoses alike violently diverge from most people’s moral intuitions. We’ve already touched on a few such scenarios above (for example, are you really morally obligated to lick the boots of someone who kicks you, just because that person is a saint to everyone other than you?). Another type of scenario involves minorities. Imagine, for instance, that 98% of the players are unfailingly nice to each other, but unfailingly cruel to the remaining 2% (who they can recognize, let’s say, by their long noses or darker skin—some trivial feature like that). Meanwhile, the put-upon 2% return the favor by being nice to each other and mean to the 98%. Who, in this scenario, is moral, and who’s immoral? The mathematical verdict of both eigenmoses and eigenjesus is unequivocal: the 98% are almost perfectly good, while the 2% are almost perfectly evil. After all, the 98% are nice to almost everyone, while the 2% are mean to those who are nice to almost everyone, and nice only to a tiny minority who are mean to almost everyone. Of course, for much of human history, this is precisely how morality worked, in many people’s minds. But I dare say it’s a result that would make moderns uncomfortable.
In summary, it seems clear to me that neither eigenmoses nor eigenjesus correctly captures our intuitions about morality, any more than Φ captures our intuitions about consciousness. But as they say, I think there’s plenty of scope here for further research: for coming up with new mathematical measures that sharpen our intuitive judgments about morality, and (if we like) testing those measures out using IPD tournaments. It also seems to me that there’s something fundamentally right about the eigenvector idea: all else being equal, we’d like to say, being nice to others is good, except that aiding and abetting evildoers is not good, and the way we can recognize the evildoers in our midst is that they’re not nice to others—except that, if the people who someone isn’t nice to are themselves evildoers, then the person might again be good, and so on. The only way to cut off the infinite regress, it seems, is to demand some sort of “reflective equilibrium” in our moral judgments, and that’s precisely what eigenmorality tries to capture. On the other hand, no such idea can ever make moral debate obsolete—if for no other reason than that we still need to decide which specific eigenmorality metric to use, and that choice is itself a moral judgment.
Scooped by Plato
Which brings me, finally, to the second new thing that’s happened this year on the eigenmorality front. Namely, Rebecca Newberger Goldstein—who’s far and away my favorite contemporary novelist—published a charming new book entitled Plato at the Googleplex: Why Philosophy Won’t Go Away. Here she imagines that Plato has reappeared in present-day America (she doesn’t bother to explain how), where he’s taught himself English and the basics of modern science, learned how to use the Internet, and otherwise gotten himself up to speed. The book recounts Plato’s dialogues with various modern interlocutors, as he volunteers to have his brain scanned, guest-writes a relationship advice column, participates in a panel discussion on child-rearing, and gets interviewed on cable news by “Roy McCoy” (a thinly veiled Bill O’Reilly). Often, Goldstein has Plato answer the moderns’ questions using direct quotes from the Timaeus, the Gorgias, the Meno, etc., which makes her Plato into a very intelligent sort of chatbot. This is a genre that’s not often seriously attempted, and that I’d love to read more of (possible subjects: Shakespeare, Galileo, Jefferson, Lincoln, Einstein, Turing…).
Anyway, my favorite episode in the book is the first, eponymous one, where Plato visits the Googleplex in Mountain View. While eating lunch in one of the many free cafeterias, Plato is cornered by a somewhat self-important, dreadlocked coder named Marcus, who tries to convince Plato that Google PageRank has finally solved the problem agonized over in the Republic, of how to define justice. By using the Internet, we can simply crowd-source the answer, Marcus declares: get millions of people to render moral judgments on every conceivable question, and also moral judgments on each other’s judgments. Then declare those judgments the most morally reliable, that are judged the most reliable by the people who are themselves the most morally reliable. The circularity, as usual, is broken by taking the principal eigenvector of the graph of moral judgments (Goldstein doesn’t have Marcus put it that way, but it’s what she means).
Not surprisingly, Plato is skeptical. Through Socratic questioning—the method he learned from the horse’s mouth—Plato manages to make Marcus realize that, in the very act of choosing which of several variants of PageRank to use in our crowd-sourced justice engine, we’ll implicitly be making moral choices already. And therefore, we can’t use PageRank, or anything like it, as the ultimate ground of morality.
Whereas I imagined that the raw data for an “eigenmorality” metric would consist of numerical measures of how nice people had been to each other, Goldstein imagines the raw data to consist of abstract moral judgments, and of judgments about judgments. Also, whereas the output of my kind of metric would be a measure of the “goodness” of each individual person, the outputs of hers would presumably be verdicts about general moral and political questions. But, much like with CLEVER versus PageRank, it’s obvious that the ideas are similar—and that I should credit Goldstein with independently discovering my nerdy 16-year-old vision, in order to put it in the mouth of a nerdy character in her story.
As I said before, I agree with Goldstein’s Plato that eigenmorality can’t serve as the ultimate ground of morality. But that’s a bit like saying that Google rank can’t serve as the ultimate ground of importance, because even just to design and evaluate their ranking algorithms, Google’s engineers must have some prior notion of “importance” to serve as a standard. That’s true, of course, but it omits to mention that Google rank is still useful—useful enough to have changed civilization in the space of a few years. Goldstein’s book has the wonderful property that even the ideas she gives to her secondary characters, the ones who serve as foils to Plato, are sometimes interesting enough to deserve book-length treatments of their own, and crowd-sourced morality strikes me as a perfect example.
In the two previous comment threads, we got into a discussion of anthropogenic climate change, and of my own preferred way to address it and related threats to our civilization’s survival, which is simply to tax every economic activity at a rate commensurate with the environmental damage that it does, and use the funds collected for cleanup, mitigation, and research into alternatives. (Obviously, such ideas are nonstarters in the current political climate of the US, but I’m not talking here about what’s feasible, only about what’s necessary.) As several commenters pointed out, my view raises an obvious question: who is to decide how much “damage” each activity causes, and thus how much it should be taxed? Of course, this is merely a special case of the more general question: who is to decide on any question of public policy whatsoever?
For the past few centuries, our main method for answering such questions—in those parts of the world where a king or dictator or Politburo doesn’t decree the answer—has been representative democracy. Democracy is, arguably, the best decision-making method that our sorry species has ever managed to put into practice, at least outside the hard sciences. But in my view, representative democracy is now failing spectacularly at possibly the single most important problem it’s ever faced: namely, that of leaving our descendants a livable planet. Even though, by and large, reasonable people mostly agree about what needs to be done—weaning ourselves off fossil fuels (especially the dirtier ones), switching to solar, wind, and nuclear, planting forests and stopping deforestation, etc.—after decades of debate we’re still taking only limping, token steps toward those goals, and in many cases we’re moving rapidly in the opposite direction. Those who, for financial, theological, or ideological reasons, deny the very existence of a problem, have proved that despite being a minority, they can push hard enough on the levers of democracy to prevent anything meaningful from happening.
So what’s the solution? To put the world under the thumb of an environmentalist dictator? Absolutely not. In all of history, I don’t think any dictatorial system has ever shown itself robust against takeover by murderous tyrants (people who probably aren’t too keen on alternative energy either). The problem, I think, is epistemological. Within physics and chemistry and climatology, the people who think anthropogenic climate change exists and is a serious problem have won the argument—but the news of their intellectual victory hasn’t yet spread to the opinion page of the Wall Street Journal, or cable news, or the US Congress, or the minds of enough people to tip the scales of history. Because our domination of the earth’s climate and biosphere is new and unfamiliar; because the evidence for rapid climate change is complicated and statistical; because the worst effects are still remote from us in time, space, or both; because the sacrifices needed to address the problem are real—for all of these reasons, the deniers have learned that they can subvert the Popperian process by which bad explanations are discarded and good explanations win. If you just repeat debunked ideas through a loud enough megaphone, it turns out, many onlookers won’t be able to tell the difference between you and the people who have genuine knowledge—or they will eventually, but not until it’s too late. If you have a few million dollars, you can even set up your own parody of the scientific process: your own phony experts, in their own phony think tanks, with their own phony publications, giving each other legitimacy by citing each other. (Of course, all this is a problem for many fields, not just climate change. Climate is special only because there, the future of life on earth might literally hinge on our ability to get epistemology right.)
Yet for all that, I’m an optimist—sort of. For it seems to me that the Internet has given us new tools with which to try to fix our collective epistemology, without giving up on a democratic society. Google, Wikipedia, Quora, and so forth have already improved our situation, if only by a little. We could improve it a lot more. Consider, for example, the following attempted definitions:
A trustworthy source of information is one that’s considered trustworthy by many sources who are themselves trustworthy (on the same topic or on closely related topics). The current scientific consensus, on any given issue, is what the trustworthy sources consider to be the consensus. A good decision-maker is someone who’s considered to be a good decision-maker by many other good decision-makers.
At first glance, the above definitions sound ludicrously circular—even Orwellian—but we now know that all that’s needed to unravel the circularity is a principal eigenvector computation on the matrix of trust. And the computation of such an eigenvector need be no more “Orwellian” than … well, Google. If enough people want it, then we have the tools today to put flesh on these definitions, to give them agency: to build a crowd-sourced deliberative democracy, one that “usually just works” in much the same way Google usually just works.
Now, would those with axes to grind try to subvert such a system the instant it went online? Certainly. For example, I assume that millions of people would rate Conservapedia as a more trustworthy source than Wikipedia—and would rate other people who had done so as, themselves, trustworthy sources, while rating as untrustworthy anyone who called Conservapedia untrustworthy. So there would arise a parallel world of trust and consensus and “expertise,” mutually-reinforcing yet nearly disjoint from the world of the real. But here’s the thing: anyone would be able to see, with the click of a mouse, the extent to which this parallel world had diverged from the real one. They’d see that there was a huge, central connected component in the trust graph—including almost all of the Nobel laureates, physicists from the US nuclear weapons labs, military planners, actuaries, other hardheaded people—who all accepted the reality of humans warming the planet, and only tiny, isolated tendrils of trust reaching from that component into the component of Rush Limbaugh and James Inhofe. The deniers and their think-tanks would be exposed to the sun; they’d lose their thin cover of legitimacy. It should go without saying that the same would happen to various charlatans on the left, and should go without saying that I’d cheer that outcome as well.
Some will object: but people who believe in pseudosciences—whether creationists or anti-vaxxers or climate change deniers—already know they’re in a minority! And far from being worried about it, they treat it as a badge of honor. They think they’re Galileo, that their belief in spite of a scientific consensus makes them heroes, while those in the giant central component of the trust graph are merely slavish followers.
I admit all this. But the point of an eigentrust system wouldn’t be to convince everyone. As long as I’m fantasizing, the point would be that, once people’s individual decisions did give rise to a giant connected trust component, the recommendations of that component could acquire the force of law. The formation of the giant component would be the signal that there’s now enough of a consensus to warrant action, despite the continuing existence of a vocal dissenting minority—that the minority has, in effect, withdrawn itself from the main conversation and retreated into a different discourse. Conversely, it’s essential to note, if there were a dissenting minority, but that minority had strong trunks of topic-relevant trust pointing toward it from the main component (for example, because the minority contained a large fraction of the experts in the relevant field), then the minority’s objections might be enough to veto action, even if it was numerically small. This is still democracy; it’s just democracy enhanced by linear algebra.
Other people will object that, while we should use the Internet to improve the democratic process, the idea we’re looking for is not eigentrust or eigenmorality but rather prediction markets. Such markets would allow us to, as my friend Robin Hanson advocates, “vote on values but bet on beliefs.” For example, a country could vote for the conditional policy that, if business-as-usual is predicted to cause sea levels to rise at least 4 meters by the year 2200, then an aggressive emissions reduction plan will be triggered, but not otherwise. But as for the prediction itself, that would be left to a futures market: a place where, unlike with voting, there’s a serious penalty for being wrong, namely losing your shirt. If the futures market assigned the prediction at least such-and-such a probability, then the policy tied to that prediction would become law.
I actually like the idea of prediction markets—I have ever since I heard about them—but I consider them limited in scope. My example above, involving the year 2200, gives a hint as to why. Prediction markets are great whenever our disagreements are over something that will be settled one way or the other, to everyone’s assent, in the near future (e.g., who will win the World Cup, or next year’s GDP). But most of our important disagreements aren’t like that: they’re over which direction society should move in, which issues to care about, which statistical indicators are even the right ones to measure a country’s health. Now, those broader questions can sometimes be settled empirically, in a sense: they can be settled by the overwhelming judgment of history, as the slavery, women’s suffrage, and fascism debates were. But that kind of empirical confirmation typically takes way too long to set up a decent betting market around it. And for the non-bettable questions, a carefully-crafted eigendemocracy really is the best system I can think of.
Again, I think Rebecca Goldstein’s Plato is completely right that such a system, were it implemented, couldn’t possibly solve the philosophical problem of finding the “ultimate ground of justice,” just like Google can’t provide us with the “ultimate ground of importance.” If nothing else, we’d still need to decide which of the many possible eigentrust metrics to use, and we couldn’t use eigentrust for that without risking an infinite regress. But just like Google, whatever its flaws, works well enough for you to use it dozens of times per day, so a crowd-sourced eigendemocracy might—just might—work well enough to save civilization.
Update (6/20): If you haven’t been following, there’s an excellent discussion in the comments, with, as I’d hoped, many commenters raising strong and pertinent objections to the eigenmorality and eigendemocracy concepts, while also proposing possible fixes. Let me now mention what I think are the most important problems with eigenmorality and eigendemocracy respectively—both of them things that had occurred to me also, but that the commenters have brought out very clearly and explicitly.
With eigenmorality, perhaps the most glaring problem is that, as I mentioned before, there’s no notion of time-ordering, or of “who started it,” in the definition that Tyler and I were using. As Luca Trevisan aptly points out in the comments, this has the consequence that eigenmorality, as it stands, is completely unable to distinguish between a crime syndicate that’s hated by the majority because of its crimes, and an equally-large ethnic minority that’s hated by the majority solely because it’s different, and that therefore hates the majority. However, unlike with mathematical theories of consciousness—where I used counterexamples to try to show that no mathematical definition of a certain kind could possibly capture our intuitions about consciousness—here the problem strikes me as much more circumscribed and bounded. It’s far from obvious to me that we can’t easily improve the definition of eigenmorality so that it does agree with most people’s moral intuition, whenever intuition renders a clear verdict, at least in the limited setting of Iterated Prisoners’ Dilemma tournaments.
Let’s see, in particular, how to solve the problem that Luca stressed. As a first pass, we could do so as follows:
A moral agent is one who only initiates defection against agents who it has good reason to believe are immoral (where, as usual, linear algebra is used to unravel the definition’s apparent circularity).
Notice that I’ve added two elements to the setup: not only time but also knowledge. If you shun someone solely because you don’t like how they look, then we’d like to say that reflects poorly on you, even if (unbeknownst to you) it turns out that the person really is an asshole. Now, several more clauses would need to be added to the above definition to flesh it out: for example, if you’ve initiated defection against an immoral person, but then the person stops being immoral, at what point do you have a moral duty to “forgive and forget”? Also, just like with the eigenmoses/eigenjesus distinction, do you have a positive duty to initiate defection against an agent who you learn is immoral, or merely no duty not to do so?
OK, so after we handle the above issues, will there still be examples that our time-sensitive, knowledge-sensitive eigenmorality definition gets badly, egregiously wrong? Maybe—I don’t know! Please let me know in the comments.
Moving on to eigendemocracy, here I think the biggest problem is one pointed out by commenter Rahul. Namely, an essential aspect of how Google is able to work so well is that people have reasons for linking to webpages other than boosting those pages’ Google rank. In other words, Google takes a link structure that already exists, independently of its ranking algorithm, and that (as the economists would put it) encodes people’s “revealed preferences,” and exploits that structure for its own purposes. Of course, now that Google is the main way many of us navigate the web, increasing Google rank has become a major reason for linking to a webpage, and an entire SEO industry has arisen to try to game the rankings. But Google still isn’t the only reason for linking, so the link structure still contains real information.
By contrast, consider an eigendemocracy, with a giant network encoding who trusts whom on what subject. If the only reason why this trust network existed was to help make political decisions, then gaming the system would probably be rampant: people could simply decide first which political outcome they wanted, then choose the “experts” such that claiming to “trust” them would do the most for their favored outcome. It follows that this system can only improve on ordinary democracy if the trust network has some other purpose, so that the participants have an actual incentive to reveal the truth about who they trust. So, how would an eigendemocracy suss out the truth about who trusts whom on which subject? I don’t have a very good answer to this, and am open to suggestions. The best idea so far is to use Facebook for this purpose, but I don’t know exactly how.
Update (6/22): Many commenters, both here and on Hacker News, interpreted me to be saying something obviously stupid: namely, that any belief identified as “the consensus” by an eigenvector analysis is therefore the morally right one. They then energetically knocked down this strawman, with the standard examples (Hitler, slavery, discrimination against gays).
Admittedly, I probably contributed to this confusion by my ill-advised decision to discuss eigenmorality and eigendemocracy in the same blog post—solely because of their mathematical similarity, and the ease with which thinking about one leads to thinking about the other. But the two are different, as are my claims about them. For the record:
- Eigenmorality: Within the stylized setting of an Iterated Prisoner’s Dilemma tournament, with side-channels allowing agents to learn who are doing what to each other, I believe it ought to be possible, by looking at who initiated rounds of defection and forgiveness, and then doing an eigenvector analysis on the result, to identify the “moral” and “immoral” agents in a way that more-or-less accords with our moral intuitions. Even if true, of course, this wouldn’t have any obvious moral implications for hot-button issues such as abortion, gun control, or climate change, which it’s far from obvious how to encode in terms of IPD tournaments.
- Eigendemocracy: By doing an eigenvector analysis, to identify who people implicitly acknowledge as the “experts” within each field, I believe that it might be possible to produce results that, on average, in practice, and in contemporary society, are better and more rational than those produced by ordinary majority-voting. Obviously, there’s no guarantee whatsoever that the results of eigendemocracy would be morally acceptable ones: if the public acknowledges as “experts” people who believe evil things (as in Nazi Germany), then eigendemocracy will produce evil results. But democracy itself suffers from a precisely analogous problem. The situation that interests me is one that’s been with us since the time of ancient Athens: one where there is a consensus among the experts about the wisest course of action, and there’s also an implicit consensus among the public that those experts are indeed the experts, but the democratic system is somehow “unable to complete the modus ponens,” because of manipulation by powerful interests and the sway of demagogues. In such cases, it seems possible to me that an eigendemocracy could improve on the results of ordinary democracy—perhaps dramatically so—while still avoiding the evils of dictatorship.
Crucially, in neither of the above bullet points, nor in their combination, is there any hint of a belief that “the will of the majority always defines what’s morally right” (if anything, there’s a belief in the opposite).
Update (7/4): While this isn’t really a surprise—I’d astonished if it weren’t the case—I’ve now learned that several people, besides me and Rebecca Goldstein, have previously written about the ideas of eigentrust and eigendemocracy. Perhaps more surprising is that one of the earlier groups—consisting of Sep Kamvar, Mario Schlosser, and Hector Garcia-Molina from Stanford—literally called the idea “EigenTrust,” when they published about it in 2003. (Note that Garcia-Molina, in a likely non-coincidence, was Larry Page and Sergey Brin’s PhD adviser.) Kamvar et al.’s intended application for EigenTrust was to determine which nodes are trustworthy in a peer-to-peer file-sharing network, rather than (say) to reinvent democracy, or to address conundrums of epistemology and ethics that have been with us since Plato. But while the scope might be more modest, the core idea is the same. (Hat tip to commenter Babak.)
As for enhancing democracy using linear algebra, it turns out that that too has already been discussed: see for example this presentation by Rob Spekkens of the Perimeter Institute, which Michael Nielsen pointed me to. (In yet another small-world phenomenon, Rob’s main interest is in quantum foundations, and in that context I’ve known him for a decade! But his interest in eigendemocracy was news to me.)
If you’re wondering whether anything in this post was original … well, so far, I haven’t learned of prior work specifically about eigenmorality (e.g., in Iterated Prisoners Dilemma tournaments), much less about eigenmoses and eigenjesus.