Dialogues on Ethical Longtermism

Is the far future most of what matters?

Two friends, M and V, sit down at a restaurant. Coincidentally, these are the same two friends who appeared in Michael Huemer’s famous book Dialogues on Ethical Vegetarianism.

M: When we were last here, you convinced me of veganism. Now, I have been vegan for (checks date when book was published) about ten years. Are you going to try to convince me of of something new this time?

V:

Specifically, I am going to try to convince you of Longtermism.

M: Oh, what’s that?

V: The basic idea of Longtermism is that making the far future go well is super important. There’s a strong form and a weak form. The weak form says: the world should be doing a lot more to make the far future go well. The strong form says: making the world go better is by far the most important global priority, so that how our actions affect the far future is much more important than how they affect present people. I accept both.

M: And why do you hold this view?

V: The argument for it is very straightforward. The future could have a lot of people. Some estimates put the number of future people there could be at 10^58, and others think it’s a lot more. If we survive for billions of years and fill the surrounding galaxies, the number of people could be truly enormous. If this is right, then in expectation, almost all of those our actions affect reside in the far future.

If future people matter, and they’re most of who we affect, then what matters most is how we affect future people. As an analogy, if there was a giant undersea civilization quintillions of times more numerous than humanity, and our actions affected them in great numbers, even threatening to extinguish them entirely, then our impacts on them would seem to be the most important impacts of our actions.

M: But can we really affect the far future? I mean, sure, maybe the far future matters a lot, but imagine people ancient Egypt trying to benefit us. They wouldn’t have been able to. Probably their guesses for how to benefit us would have been weird stuff like “pray to Ra for the continuation of future humanity.” As more and more time passes, it becomes harder and harder to predict the impacts of your actions. So over long time scales, our ability to predict decays to almost nothing.

V: I think that’s basically right for lots of effects of our actions. In many cases, you can’t really predict how an action will affect the future. But there are some cases where you can. These are the most important impacts of our actions.

A persistent state is one that, if entered, will stick around. Extinction is a persistent state. If we go extinct, we’ll stay extinct. So by influencing the probability of extinction, we can affect the far future predictably. Other persistent states include: lock in (where we cement our values in stone for ever), extreme power concentration, and institutional development of, for example, space.

In fact, lots of experts think that AI might produce lots of rapid economic growth, and give us the ability to preserve current institutions forever. So affecting the trajectory of AI—ensuring the AIs shaping the future have good values and care about everyone’s welfare—could have enormous and lasting effects. If other actions are comparable, it’s because they benefit far future people.

M: None of that sounds crazy, I guess, but is that obviously more effective than other charitable opportunities.

V: Yes! If a trillion dollars would reduce existential threats by one in a billion, and you think the earlier 10^58 number if off by 10 orders of magnitude, then the expected number of life years added by a single dollar is 10^27. That’s a ridiculous amount—far more than the total number of people on Earth. It would be extremely surprising if there weren’t any spending opportunities that could reduce ex risks by more than one in a billion for a trillion dollars!

M: Sure, but actions could backfire as well. Maybe you’ll try to regulate AI and actually end up making things worse. History is replete with examples of people trying to do good and it backfiring. So perhaps we just don’t have enough information to reliably improve the future.

V: Sure, it’s always possible our actions will backfire. But note: that doesn’t just apply to actions to benefit the far future. It applies equally to actions to benefit present people. So if you think that some actions done for the sake of present people are good in expectation, then the same should be true of future people.

Let’s be a bit more concrete. Suppose you get a career working to stop deadly pandemics. Might that backfire? Sure, I guess. But while it might have bad effects you didn’t think of, it might also have good effects that you didn’t think of. Having good effects you didn’t think of seems, if anything, more likely. So then I don’t think this should undermine your confidence in this being a good career choice.

I don’t think you can really be completely in the dark about whether any action benefits the far future. If you were, then you’d think that working to cause a pandemic was just as likely to benefit the future as not to. That doesn’t seem right. So while we should remain cautious and think seriously about backfire risks, they aren’t reasons for total inaction. And if you think causing deadly pandemics is bad for the future in expectation, you should also think preventing deadly pandemics is good for the future in expectation.

M: That might be right. But doesn’t this assume utilitarianism? What if I don’t care about maximizing total welfare?

V: No, it doesn’t assume utilitarianism. It just assumes that the interests of the overwhelming majority of those affected by your actions matter.

M: And doesn’t it assume moral realism? Which I reject completely—people say it’s intuitive, but it’s certainly not intuitive to me. And whether it’s intuitive to most people is an empirical question.

V: No, why would it assume that? It’s just a standard moral claim. You don’t have to think morality is objective to make moral claims sometimes.

M: And doesn’t it assume disjunctivism about perception?

V: Um, no, why would you think that?

M: And doesn’t it assume that the French renaissance was highly overrated as a period of historical inquiry.

V: No.

M: Okay, okay, but I think I might have something it does assume. Doesn’t it assume that creating happy people is valuable? I reject that view. I was reading Jan Narveson, and he said “make people happy, not happy people.” That really resonated with me. The slogan just sounded so true.

V: And why would it assume that?

M: Well, you kept talking about how many future people there might be—about how extinction might decrease the number of future people. But if I adopt the person-affecting view, then I don’t care about increasing the number of happy people.

V: Remember what Longtermism says: making the future better is really important. It doesn’t, on its own, say anything about which things make the future better. So even if you think filling the future with happy people doesn’t matter, you can still be a Longtermist. Then you should think that preventing the future from going badly and preventing future suffering is important.

M: And how might we do that?

V: A number of ways. You might try to fight for moral progress, because moral errors might lead to lots of suffering. You could try to increase the representation of digital minds or work to give AIs better values. Alternatively, you can try to prevent extreme power concentration.

On top of this, the person-affecting view is very implausible—for more on this, see a really nice blog post by the dashingly handsome blogger Bentham’s Bulldog, rivaled in handsomeness only by his cleverness and most of all, his humility.

M: Okay but what about this. As John Broome has shown, the person-affecting view holds that creating people is greedy. This means it can soak up value in either direction. So if you create a happy person and make someone slightly worse off, that neither makes the world better nor worse. Similarly, if you create a happy person and make someone slightly better off, that makes the world neither better nor worse.

So by changing the distribution of future people, won’t this soak up all the value of Longtermist interventions?

V: Well, all of our actions change the distribution of future people. So really this is a fully general objection against any actions being for the best.

M: But isn’t this a very strange moral argument to make? It’s very counterintuitive that the far future matters this much. Just as we can toss out an ethical theory if it says slavery is okay, can’t we toss out one for saying that all of our actions’ impacts on future people are a rounding error? On this view, if the world could voluntarily take a 90% reduction in present welfare, all to improve the welfare of people alive more than a million years in the future, we should. That’s very counterintuitive.

V: I agree it’s a bit surprising. But sometimes moral claims are surprising because the world is surprising, not because the ethical claim is surprising.

Remember when we talked about meat-eating. You said it was counterintuitive that factory farming was the worst thing in the world. But then we agreed that wasn’t really counterintuitive. When we broke down the core argument, it was:

1. If something is the cause of more pain and suffering in a few years than all the suffering in human history, then it’s one of the worst things in the world.

2. Factory farming is the cause of more pain and suffering in a few years than all the suffering in human history.

3. So factory farming is one of the worst things in the world.

It’s not premise 1 that’s counterintuitive but premise 2. If a surprising moral claim arises because the world is surprising, that’s no reason to reject the ethical theory underpinning it. When the world is surprising, our judgments about it should be as well. Let’s do the same for the case for strong Longtermism. The core argument is:

1. For every present person our actions affect, there are in expectation trillions of future people that our actions affect.

2. If an action affects trillions of times more expected future people than present people, then the main thing that determines whether the action should be performed is how it affects future people.

3. Therefore, the main thing that determines whether actions should be performed is how they affect future people.

The surprisingness resides in premise 1, not premise 2. So this is no reason to reject strong Longtermism. When the world is weird, the ethical claims that apply to it will be weird too.

M: You’ve been talking about how many future people there will be. But the estimates you’ve been given have been assuming that there could be a lot of digital people. I reject that! I think consciousness is substrate dependent—no life, no mind, no person.

V: First of all, the case for strong Longtermism doesn’t depend at all on whether there can be digital people. The future can be really big even if there can’t be lots of happy digital minds. Second, you should have some serious uncertainty about whether there can be digital minds. Many of the best philosophers in the world think there can be. Even on standard views according to which consciousness is substrate dependent, there can still be digital minds if they’re configured correctly. But even if you think there’s only a 1% chance that there might be digital people, there are still enormous numbers of them in expectation.

M: But doesn’t this all assume Fanaticism? Your odds of making any major difference to the far future—preventing extinction, say—are very low. So then isn’t this a Pascal’s mugging? It’s a tiny chance of a big impact.

V: No, being a non-fanatic isn’t any help, and this isn’t a Pascal’s mugging. What alternative to fanaticism do you have in mind?

M: I think you should discount low risks. If someone comes to you and offers you infinite utility, and you’re 99.99999999999999999999% they’re lying, then you should value their offer at nothing even if its expected value is high.

V: Do you mean you simply ignore every event with low probability?

M: Yes.

V: That’s very implausible

M: Why?

V: Suppose that someone has a quintillion ways to torture you. They’ll pick a random one. Should you say “well, the odds of every particular one being picked are one in a quintillion, so this is no big deal—every specific scenario gets counted for nothing?”

M: Maybe the threshold is below one in a quintillion.

V: Whatever the threshold is, just change the number to that. Also, obviously the odds of Longtermist interventions making the far future a lot better are way more than one in a quintillion.

M: Hmm, fair point, I had never thought of that and have never thought about philosophy of risk at all. So then I guess the better view is the one adopted by the academics called “tail discounting,” rather than naive Nicolausian discounting that’s fallen significantly out of favor. Tail discounting involves discounting the best and worst slices of probability space. So when you’re tail discounting, you ignore the best and worst .000000001% of scenarios and only focus on the middling ones.

V: But Longtermist interventions nudge things outside of the best .000000001% of scenarios. So tail discounting is no help.

M: Hmm, good point, I’ll have to think more on that one. In light of this, are there any other considerations that undermine in serious ways the risk discounting objection to Longtermism?

V: Indeed there are. But they don’t lend themselves great to dialogue format, so I’ll just point you in the direction of section six of this article by dashingly handsome blogger Bentham’s Bulldog—whose good looks are matched only by his unwillingness to put self-praise in the mouth of dialogue characters.

M: Okay, now that all of my objections have been refuted, I am convinced! I’m now a strong Longtermist. But what should I do about it?

V: Well, a lot of impactful careers are advertised on the 80,000 hours job board! They also have more general career advising—and see here for some high-impact Longtermist projects. So go there and try to get a high-impact career that makes the world better.

And thus it was that M became a polyamorous vegan alignment researcher in a bay area polycule.