Epistemology in the age of AI

I bring you good news of great joy

Today, I depart from my customary pessimism and bring you good news. Superintelligence (if it occurs) is likely to be ethical.

When I speak with people in fields like science and technology, I often come away very impressed by their knowledge of artificial intelligence. So much so that I don’t believe that I know enough to have a useful opinion on the likely future trajectory of AI. But when it comes to epistemology, I am greatly underwhelmed by what I encounter, even from people with very high IQs. This raises an interesting question: Do epistemologists have anything useful to contribute to the debate over AI?

From the outside, epistemology seems like a field that has failed to produce a coherent model of knowledge. That’s partly because the concepts are so hard to pin down. Terms like knowledge, belief, science, truth, fact, theory, opinion, objective, subjective, and reality are hard to define. When the terms are defined, it is often in terms of other equally vague terms. I see lots of circular definitions. And it seems clear to me that these terms mean different things to different people.

In part 2, I plan to argue that future AIs will be highly ethical, but to do so I need to explain my views on epistemology.

Part 1: Aesthetic and ethical knowledge

Many people that I know in “hard” fields like science, technology, economics and finance have a view of epistemology that creates a sort of dichotomy between two realms, which I’ll call knowledge and belief:

Knowledge: Science, fact, reality, truth, objective

Belief: Ethics, aesthetics, opinion, theory, subjective

In my view, this sort of strict dichotomy is really, really bad epistemology. For instance, I see people denying that the concept of “knowledge” applies to the field of art. They say things like:

It’s just a matter of opinion.

Art should be pretty.

What counts is popularity.

Modern art and architecture is pretentious rubbish.

Let’s start with popularity. If artistic merit is nothing more than current popularity, then how can there be any “testable implications”, such as a prediction of future popularity? But don’t we already know that 100 years from now, Beethoven’s 9th symphony will likely outsell today’s #1 pop song?

In 1985, I made my first trip to the Prado, and was blown away by Las Meninas. For 20 minutes the painting washed over me in waves, as I gradually absorbed more and more. Had I been primed by its reputation? Maybe, but not so for many other great paintings in that museum, which I only later discovered were highly esteemed. I had never even heard of Zurbarán. Compare this Zurbarán to a cute Murillo in the same museum. Which is prettier? Which is better? Who says so?

There is a difference between prettiness and beauty. Yes, the words overlap. But the connotation of one is simple, whereas the other is rich and complex. It’s analogous to difference between a sweet cherry soda pop and a complex red wine.

The art world is not a democracy. Art reputation is based on “public opinion”, but severely weighted by expertise. But isn’t expertise also subjective? What is not subjective is intense interest. Who spends an hour at the Prado, and who spends all day? And it is not subjective to say that a reputation endures for centuries.

[As an aside, if you have only 30 minutes for the Louvre, do not see the Mona Lisa. If you are not interested in art, see Death of Sardanapalus. If you insist on a work by Leonardo, see the Virgin of the Rocks. If you want to see great art, check out the best French old masters (Poussin, Chardin, etc.)]

Here’s what art skeptics don’t get. They divide the world into phony elitists who pretend to like modern art, and sincere regular people who favor pretty realistic art. But that’s factually incorrect. To a large extent, the people who go to the MOMA are the same people that go to the Met. The actual distinction that matters is between those with an intense interest in art, all forms of art, and those with only a passing interest in art—some decoration to put on the wall.

The world of art is controlled by the people with an intense interest in the field. They determine what gets put in museums. They determine which painters become famous. They determine how the art history textbooks are written. And for the most part, they like both old masters and modern art. Average people? They like impressionism. If they are less educated, they like Thomas Kinkade. They like pretty art.

There’s no shame in having mediocre taste in art—I have mediocre taste in music and poetry. If I repeatedly listen to a complex piece of music, it gets better each time. My ear is being “educated” by repeat listening. But I plateau well below the level of a true music lover, due to my sub-optimal brain wiring.

I implore you to avoid hasty judgments. Don’t assume that just because you don’t see it, it’s not there. Art is not fake. There really is something there, beyond a pretty picture. Great art is rich and strange and complex and beautiful. Just as Beethoven’s 9^th is much more than just a pretty song like Ode to Joy.

Nature also produces aesthetic beauty. And that beauty cannot be reduced to a set of fixed rules. The thing that makes the Matterhorn beautiful is not the same thing that makes Mt. Fuji beautiful. And neither contains the thing that makes Uluru (Ayer’s Rock) beautiful. The reason that Niagara Falls is beautiful is completely different from the reason that Angel Falls is beautiful.

And don’t confuse style with greatness. Styles come in and out of fashion. That doesn’t mean the previous style was ”wrong”. People always seem to want something new. But greatness within a style tends to endure, once it is recognized. (It may take hundreds of years to be recognized, but once it is . . .) Ruben’s paintings are clearly out of style, but just as clearly are still great.

Art is not the only field where knowledge affects one’s aesthetic appreciation. Mathematicians often describe notable theorems or proofs as “elegant” or “beautiful”. You might argue that mathematics is a “rigorous” field, subject to “proof”. Art is not. I agree. But the beauty of a mathematical proof is every bit as subjective as the beauty of a painting.

If 10 esteemed mathematicians tells me that Cantor’s diagonalization argument is a beautiful proof of the existence of uncountably infinite sets, I’ll take them at their word. I don’t care if 100 plumbers say they don’t see any beauty in the theorem. Call me an elitist, that doesn’t bother me. Mathematicians get to determine what math theories are “great”, not plumbers. And art experts get to determine what art is considered great.

Don’t think so? Imagine that an average plumber is on a TV game show, and asked to pick a great artist from a list of names, including Bouguereau and Picasso. Which name does he pick? Which artist does he likely prefer? The expert definition of greatness is implicitly accepted even by people that don’t agree with expert opinion.

You might argue that mathematicians have knowledge that gives them a better ability to judge the beauty of a proof, whereas when looking at a painting there is nothing to know. It’s just an image that appears on your retina, as it appears on anyone else’s retina. But that’s not how vision works, the image is processed by the brain in a very complex fashion. You don’t see the world directly; your brain “thinks” an image into consciousness, along with all its emotional and intellectual connotations.

After I took some college courses in art history (and architecture history), I began to see aspects to art and architecture to which I had previously been oblivious. There really is such a thing as aesthetic knowledge, just as there is scientific knowledge and ethical knowledge. Indeed I’d argue that all knowledge fits into one of those three boxes: ethics, science and aesthetics—roughly the good, the true and the beautiful. Is there any other type of knowledge?

Obviously, most of humanity could not give a damn about a painting like Las Meninas. They don’t see what I see, or what I think I see. That’s fine. But to convince me that I’m wrong about aesthetic knowledge, you’d have to convince me that I’m hallucinating when I look at that painting. Even worse, you’d have to convince me that lots of other art experts with much more aesthetic knowledge than I have experience a similar hallucination. That’s going to be a really hard sell.

I’ve seen people in the so-called hard sciences dismiss aesthetic and ethical knowledge as “mere opinion”. I suppose from a certain perspective everything is mere opinion—even scientific laws. But clearly the negative connotation of “mere opinion” doesn’t do justice to the rigor and usefulness of science. And while science has a greater level of rigor, all three areas have propositions that we are highly confident about, and other propositions about which we are much more uncertain.

There are times when even making a sincere effort doesn’t allow one to grasp an aesthetic concept. But the same is true in science. On several occasions, I’ve read explanations of things like time dilation and superposition, but I cannot honestly say I truly understand special relativity or quantum mechanics. I took a short course on musical appreciation, but still failed to understand much of the music they discussed.

Fields such as aesthetics and ethics have much more real knowledge than many people in the hard sciences assume. Experts in those “soft” fields know lots of things, and the things they know are often very useful.

Consider the following exchange:

Henry [Oliver] A lot of people think that Shakespeare is overrated. We only read him because it's a status game. We've internalized these snobbish values. We see this stated a lot. What's your response to these people?

Tyler [Cowen] Well, I feel sorry for them. But look, there's plenty of things I can't understand. I just told you if I go to see the plays, I'm completely lost. I know the fault is mine, so to speak. I don't blame Shakespeare or the production, at least not necessarily. Those are people who are in a similar position, but somehow don't have enough metarationality to realize the fault is on them. I think that's sad. But there's other great stuff they can do and probably they're doing it. That's fine.

And here’s Bob Dylan, channeling GK Chesterton:

Don’t criticize what you can’t understand.

Ethical knowledge is also real. South Korea is not better than North Korea because it has better science. The North has good scientists; they even created a nuclear weapon. The South is better because its government has better ethics—it’s closer to classical liberalism.

In my view, the invention of the novel did more to create the industrial revolution than the invention of the steam engine. The novel (and more broadly the narrative arts and religious texts) led to empathy for “the other”, which is was what led to classical liberalism, which allowed the Dutch, then the English, and then the Americans to transform the entire world.

Part 2: What does any of this have to do with AI?

If there’s no such thing as ethical knowledge, if it’s all “just opinions”, then we cannot have any confidence in our ability to predict the ethics of a future super-intelligent AI (ASI), at least beyond what we’ve programmed in. And lots of doomers worry that we’ll fail in any attempt to program in good values.

But suppose there is such a thing as ethical knowledge. In that case, you’d expect AIs to become more ethical as they become smarter. An ASI smart enough to fake alignment would likely also be smart enough to understand that this is not a good idea.

At this point people often raise the objection that there are smart people that are unethical. That’s true, but it also seems true that, on average, smarter people are more ethical. Perhaps not so much in terms of how they deal with family and friends, rather how they deal with strangers. And that’s the sort of ethics that we really need in an ASI. Smarter people are less likely to exhibit bigotry against the other, against different races, religions, ethnicities, sexual preferences, genders, and even different species.

In my view, the biggest danger from an ASI is that the ideal universe from a utilitarian perspective is not in some sense what we want. To take an obvious example, it’s conceivable that replacing the human race with ten times as many conscious robots would boost aggregate utility. Especially given that the ASI “gods” that produced these robots could create a happier set of minds than what the blind forces of evolution have generated, as evolution seemed to favor the “stick” of pain over the “carrot” of pleasure.

From this perspective, the biggest danger is not that ASIs will make things worse, rather the risk is that they’ll make global utility higher in a world where humans have no place.

Part 3: A bit of armchair epistemology

I’ll conclude with a few observations about epistemology. Let’s start with objective truth. What does that really mean?

1. Matches reality?

2. Known with certainty?

3. Not conditional on the observer’s frame of reference?

Those are three different concepts. Consider this quote from the Oliver/Cowen interview:

Henry How objectively can we talk about art?

Tyler I think that becomes a discussion about words rather than about art. I would say I believe in the objective when it comes to aesthetics, but simply because we have no real choice not to. People actually, to some extent, trust their aesthetic judgments, so why not admit that you do and then fight about them? Trying to interject some form of extreme relativism, I think it's just playing a game. It's not really useful. Now, is art truly objective in the final metaphysical sense, in the final theory of the universe? I'm not sure that question has an answer or is even well-formulated, but I would just say let's just be objectivists when it comes to art. Why not?

This is similar to the view of Richard Rorty:

Thus the dispute between Engel and myself does not bear on the question of knowing whether there is something that we call objective knowledge. That we use this term is obvious. What divides us is the question whether we should say that certain areas of inquiry attain such knowledge, whereas others unfortunately cannot.

My view is that terms like objective and subjective do more to confuse than enlighten. As Tyler says, it’s often a “discussion about words”. But if we insist on using the term in science, let’s also use it in ethics and aesthetics. And is even science completely objective? Consider the following quote from Quanta magazine:

Accepting this conditional definition of entropy has required a rethinking of the fundamental purpose of science. It implies that physics more accurately describes individual experience than some objective reality. In this way, entropy has been swept up in the larger trend of scientists realizing that many physical quantities make sense only in relation to an observer. (Even time itself was rendered relative by Einstein’s theory of relativity.) “Physicists don’t like subjectivity — they’re allergic to it,” said Anthony Aguirre(opens a new tab), a physicist at the University of California, Santa Cruz. “But there is no absolute — that’s always been an illusion.” . . .

Most people are content to view taste in certain foods as subjective. John’s claim that he prefers blueberries is not falsified by Jane saying that she likes raspberries best. Their tongues and/or brains might be wired differently. They both might be right—for themselves. On the other hand, it makes no sense to claim that 2+2 = 4 is true for Jane but not for John.

Many cases fall between these two extremes. Tyler would probably say that James Joyce’s Ulysses is objectively better than a Ludlum novel, even if most people would prefer to read the latter. How should we think about that claim?

In my view, taste in novels is partly objective and partly subjective; at least in the sense Tyler is using the term objective. Through education, people can gain a great appreciation of Ulysses. In addition, Ulysses is more likely to be read 100 years from now than is a random spy novel. And most experts prefer Ulysses. All three facts are relevant to the claim that artistic merit is partly objective. Like scientific knowledge, there is consideration of evidence, appeal to authority, and the test of time. Again, here’s Rorty:

I would maintain that a person who says “that belief is justified, but perhaps not true” should be taken to be distinguishing not between something human and something nonhuman but rather between two situations in which human beings may find themselves: the present situation, in which the belief appears to be justified, and a hypothetical situation in the future, where it will no longer appear justified.

On the other hand, the raspberry/blueberry distinction based on taste suggests that an art form like the novel is evaluated using both subjective and objective criteria. For instance, I suspect that some people (like me!) have brains wired in such a way that it is difficult to appreciate novels looking at complex social interactions with dozens of important characters (both men and women), whereas they are more open to novels about loners who travel through the world and ruminate on the meaning of life. Other people have brains wired in such a way that subtle social dynamics are easy to read. Thus between two novelists that are roughly equally rated by experts, there will be disagreement over ranking based on differing taste. Neither preference is necessarily wrong.

STEM people often get excessively stuck on the question of realism vs. anti-realism. Are ethical beliefs facts about the universe, or arbitrary human constructs? Rorty argues that we need to grow up—we have no access to a God’s eye view of things. He suggests that we need to get on with discovering scientific knowledge, ethical knowledge and aesthetic knowledge, and not obsess over distinctions that have no utility:

For the fundamental thesis of pragmatism is William James’s assertion that if a debate has no practical significance, then it has no philosophical significance.

So my objection to the “realism vs. anti-realism” debate is not that the debaters are deploying sentences that are devoid of meaning, nor that they are using terms that do not designate substantial properties. Rather it is that the resolution of these debates will have no bearing on practice. I view debates of this sort as examples of sterile scholasticism.

If the ASI optimists (or are they pessimists?) are correct, then perhaps we may soon have a God’s eye view of truth. Rorty’s long run “time will tell” might happen much sooner than any of us imagined just a few years ago.

Suppose that by 2027, the best software engineers are AIs that already exceed their human counterparts, and that AI progress then increases exponentially. By 2030, we have superintelligence (ASI). What then?

I presume we will quickly learn whether the universe is mostly dark energy, or whether that concept is a fudge factor for flaws in our cosmological model. Relativity and quantum mechanics will be unified. Cancer will be cured. Those changes are not hard to imagine.

What STEM types often overlook is that we’ll also have explosive progress in ethics. Longstanding debates over abortion, the death penalty and teenage hormone therapy will be quickly resolved. We will no longer need a Supreme Court to decide these issues.

In my field (macroeconomics), the ASI will create a model that will be so dazzling, so obviously true, that Paul Krugman, John Cochrane and myself will all see its merit, and discard our current (conflicting) views. Even MMTers will acquiesce.

Do I think all of this will actually happen? I don’t know. I do accept there is such a thing as ethical knowledge, and thus I believe that if super-intelligence actually occurs, it will be analogous to a religious event, like God coming down from heaven and telling us all the OBJECTIVE TRUTH.

And my outside view is that the Silicon Valley types know much more about this than I do, so I should defer to their opinion.

My inside view is that this will not happen; it will be much harder to produce superintelligence than we now imagine. But that’s based on nothing more than my gut instinct, and no one should put any weight on that opinion.

Not long ago, I didn’t expect to live long enough to see these questions resolved. Now it seems moderately likely that I will—I’ll only be 75 years old in 2030, and my current life expectancy is probably into the low 80s. Curiosity about all of this gives me a motivation to keep living a few more years.

PS. After writing this post, I can across the following in an excellent Scott Alexander post:

I don’t think anyone is, deep down, a based post-Christian vitalist. It’s fun to LARP as the Nietzschean superman, but ask Raskolnikov how far that gets you. I think we all have the same basic moral impulses, and that for most people - including most people who deny it - those potentially include caring about poor people you’ll never meet, suffering in far-off countries.

I think he’s right, and I believe this will also apply to future ASIs.

PPS. A study of aesthetics shows some surprising connections. Consider the stereotypical East Asian martial arts film, with characters doing impossibly difficult leaps while fighting (something like Crouching Tiger). You might assume that this is a modern invention. But the following Yoshitoshi woodblock triptych from the 19th century shows that the motif has deep roots in Asian culture: