Against Induction From "Silly People Thought X"

It doesn't matter how many confused people in the past thought things like what you think now

Sarah

over at the blog longer ramblings has a nice article called Are AI safetyists crying wolf? People often claim that artificial intelligence fears are just the newest flavor of tech panic—that people had the same fears about trains, nuclear power, airplanes, and every other kind of technology. Thus, given that panics over tech are ubiquitous and consistently wrong, we shouldn’t take the latest tech panic all that seriously. If doomsday prophets predict the end of the world ten times, and are always wrong, you shouldn’t take their 11th prediction of doom all that seriously.

As Sarah notes, claims that people predict every new kind of technology will end the world are greatly exaggerated. It really isn’t true that people thought trains would cause the end of the world. And while some people had fears about new technologies, these were often not widespread, and these fears were often reasonable. Fossil fuels have caused pollution to kill hundreds of millions of people throughout history; nuclear weapons made the annihilation of developed life possible for the first time in history; cell-phones have likely been responsible for a severe mental health crisis. It’s only reasonable to complain about overwrought fears from new technology if you can demonstrate that new technology is feared too much. However, demonstrating that would require a detailed historical survey, rather than just cherry-picking a few examples of people making silly claims about new technology.

Thus, the argument that we shouldn’t take AI risk seriously because people have repeatedly thought new technology would end the world, is wrong on empirical grounds. But I think it has a deeper problem: even if it were right, it wouldn’t undermine the case for AI risk.

Suppose that lots of people, for most of history, had thought that evolution by natural selection explained various things that it clearly does not. For instance, such people would claim that evolution explains the existence of rocks, clouds, and the planet Jupiter. Their arguments were poor and published mostly in tabloid journalist outlets, rather than being made by serious scientists.

Suddenly, Darwin comes along and shows that lots of facts about biology can be nicely explained by evolution by natural selection. It would not be reasonable to deny the significance of Darwin’s finding on the grounds that stupid people have been crying wolf for a long time. If you find smart, reasonable people making reasonable arguments at some time, it doesn’t matter how many people have been making crazy arguments before.

Now, it would be one thing if the cases in the past were strict parallels of AI risk. If, for instance, several Nobel prize winners, as well as a sizeable share of the leading experts on train safety, predicted that trains world end the world, and had been wrong, you might suspect that a similar thing is going wrong regarding AI risk. If over and over again, many leading experts produced carefully argued reports arguing that some new technology would end the world, and they turned out wrong, that would be decent evidence against AI doom.

But nothing remotely like that has happened. No one produced carefully argued briefs claiming trains would end the world, as there’s no argument for trains ending the world that’s not obviously ridiculous. Those working on trains didn’t think, on average, that there was about a 10% chance that trains would end the world. The leading experts on trains didn’t think that. No one predicted trains ending the world, and serious fears about them were tabloid newspapers stoking ridiculous fears rather than anything serious.

My basic point in a sentence: arguments of the form “people not thinking very carefully thought things like X for ridiculous reasons,” doesn’t give you much evidence against X, if smart people in the present think X, and their reasons aren’t obviously ridiculous. If the kind of consensus around X in the present isn’t anything like the previous support for the things like X, and the arguments are far better, then they’re not an analogous reference class.

Now that we’ve gotten clear on this error, we can see what’s wrong not merely with arguments against fearing the upcoming robot gods, but with arguments against God most high (אֵל עֶלְיוֹן‎). A common claim is that arguments for God commit the god of the gaps fallacy. The basic idea is roughly that God is used only to fill the gaps in our knowledge—that we previously used gods to explain lightning, movement of the waves, and so on. Once we learned more, we came to see that those were explained by natural processes. Thus, it is claimed, arguments like fine-tuning are just the newest edition in a long historical process of vacuously using gods to explain things that we don’t understand. As those beliefs were wrong in the past, so too will modern arguments for God be shown to be wrong.

Now, like crime in multistory buildings, this is wrong on a number of levels.

First of all, the historical claim is wrong. Theists have not historically used God to fill every gap in our understanding. As Aron Wall pointed out, until recently, we didn’t know why holes in Swiss cheese formed. Nonetheless, nobody thought God did that directly. As Aron writes:

To the best of my knowledge, no Christian apologist has ever made the following argument: 1) Science cannot explain high temperature superconductivity [a puzzling phenomenon in condensed matter physics], 2) therefore an intelligent designer must have caused it, 3) therefore God exists. The reason is that it is obvious in this case that there should exist in principle an ordinary scientific explanation for this phenomenon. Superconductors involve complicated, messy physics and there is no particularly good reason to be surprised that we don't understand them fully yet.

While some polytheists have argued that gods were behind natural phenomena—e.g. the Greeks thought Zeus was behind lightning—this wasn’t generally an argument for Zeus. Contrary to the claims of those who make this charge, there is not a long tradition of people finding random natural events puzzling and saying that God must be the explanation of them. Not even the smart polytheists made these arguments—Aristotle and Plato didn’t make these arguments, for instance. This will be shocking to some, but if you get your view of history from Christopher Hitchens’ polemics, where his sources were pulled directly out of his ass, you will get a misleading picture! In fact, historically monotheists have been behind scientific advances, believing a good God would make an intelligible world that unfolds according to natural laws.

Secondly, there’s nothing fallacious about using God to explain stuff. Anytime you do abductive reasoning, you’re invoking some entity to fill a gap in our understanding. It’s not fallacious to think that some evidence from evolution is that it fills gaps in our understanding—explaining, for instance, why gorillas have the thumbs they do and why there are transitional fossils. It’s not fallacious to think Thales existed, on the grounds this fills a “gap” in our understanding—namely, all the writing about Thales!

So long as we have some reason to expect some phenomenon if God exists, and we have reason not to expect it if he doesn’t, then it’s evidence for God. This isn’t a fallacy, it’s just Bayes theorem. Some fact is evidence for a hypothesis if it’s likelier if the hypothesis is true than if it’s false. If we have no clue why something would be the case if there was no God, and yet have active reason to expect it if there is a God, then straightforward probability theory implies this is evidence for God.

But fine, let’s pretend that the historical claim is right. Let’s suppose that theists have, since the dawn of time, pointed to rocks, stars, mysteriously eaten sandwiches, your furniture not being where you left it, swiss cheese, and oceans, and claimed these things were best explained by God. Should we then distrust abductive arguments for God—from consciousness, fine-tuning, and so on?

No! Here’s one difference between the argument for God from waves and the argument from fine-tuning: the first one is obviously stupid and the second one isn’t. There’s no particular reason God would like waves and there’s no particular reason waves are hugely improbable if God doesn’t exist. The argument for God from waves is obviously idiotic (except, insofar as the point is simply that waves require cosmological fine-tuning, in which case it’s just a disguised fine-tuning argument).

In contrast, cosmological fine-tuning is seriously puzzling. Top physicists generally regard it as a great mystery. It’s supported by powerful scientific evidence, unlike the argument from waves. You might ultimately reject it, but it’s not obviously stupid in the same way the argument from waves is.

Following the earlier principle, we shouldn’t do induction of the form “people who didn’t think very carefully thought ridiculous things similar to X, therefore we should reject X,” when X is not ridiculous and isn’t thought primarily by unthinking people. Even many atheist philosophers think fine-tuning is a good argument—not so with the argument from waves or swiss cheese. Thus, even if those making the God of the gaps charge were right about history, this would be almost entirely irrelevant. One’s inductive reference class when treating serious and clever arguments shouldn’t be filled almost exclusively with stupid and terrible arguments.

If we’re going to exclude the obviously terrible arguments from our reference class, then the historical record looks a lot murkier regarding both God and AI. Regarding AI risk, there has only been one previous technology that serious people thought could end the world for reasons that weren’t obviously ridiculous: nuclear weapons. But fears of nuclear weapons weren’t overblown—they really are a major risk to the world. Thus, the track record of fears of new technology backed by serious argument isn’t bad.

Regarding God, a similar pattern holds. There was one argument for God’s existence that was good, and has been debunked by scientific advances: the argument from biological life. For much of history, it was reasonable to look at the diversity of biological life as strong evidence for a creator. However, the discovery of evolution by natural selection (mostly) rubbished that argument.

But in its place, many new very powerful arguments have sprung up. I recently made a tier list of arguments for God. None of the best arguments was more 50 years old. As we make scientific—and especially philosophical discoveries—the case for theism looks better and better. Philosophical investigation into anthropics turned up the anthropic argument. Investigation into philosophy of mind turned up psychophysical harmony. Investigation into fundamental physics turned up fine-tuning. While most of the objections to theism—like the problems of evil and hiddenness—are quite old, the case for theism is like a fine wine: it gets better with age. If we had only arguments that had been made prior to 1980, I would be an atheist—and quite a confident one!

Thus, there’s nothing like a clear inductive trend that undermines either the case for God’s existence or the case for AI risk. It’s only if you fill the reference class with a bunch of bullshit that you get a trend against taking those things seriously—but this is a general feature of filling any reference class with a bunch of bullshit. You should simply not do that!