Anthropic reasoning 1: I am, therefore I think.

Should you use the mere fact of your own existence as a significant piece of information in your reasoning? Or is it, in and of itself, without any actual specifics, completely vacuous or even incoherent? To explain what I mean, here's a concrete example. 


Does the mere fact that we, denizens of the planet Earth, exist give us reason to think that life on other planets is not too rare? Suppose, for simplicity, that microbiologists and chemists figured out all the steps necessary for life to evolve on any planet, and how likely each step is, except for one mystery step. The mystery step is: for life to evolve, molecules X and Y must be able to combine and form XY. The scientists conducted extensive calculations and arrived at only two possibilities:

  • M = they combine very easily, as a result many planets per galaxy develop life.
  • S = they combine very rarely, as a result only a single planet per galaxy develops life, on average.
Suppose from the perspective of chemistry the two possibilities are equally likely. The million dollar question is: does our existence give us reason to lean in favor of life being pretty common in the galaxy? In other words, should we believe that M is probably true?

You might think the answer is completely obvious. Except whether you answer yes or no, there are lots of smart people who think the answer is completely obviously the opposite! Philosophers have been arguing about scenarios like this for decades with no end in sight. Besides being a really cool mind-breaking puzzle, the question is not purely academic, it has profound scientific and even public policy implications. 

What's the answer?

I will put my cards on the table: I claim our existence is a crucial datum. In the above story we definitely need to favor M. The basic reason for this can be summarized in one sentence like this:

We are more likely to be alive today if there were more "attempts" in which we could be made.

This "bumper sticker" is not really an argument in itself, it's more like a hint at an argument. But I think there are several solid arguments for my position, and I want to give one here, one which I think will simultaneously clarify what it even means to use our existence as a datum in the first place. First, I'll just give a quick list of other arguments:

  • The ensemble analysis from my recent post.
  • The opposite view entails a bizarre ability to predict the future, such as the doomsday argument.
  • The opposite view requires vague arbitrary rules about what counts as an observer, or "me".
  • My view is in accord with TER, which says it's ok to use all available data.
  • My view leads to intuitively correct results in many situations.
  • In the opposite view, completely irrelevant information affects the answer.
So far I have discussed the first of these arguments, the ensemble analysis, which I think is probably the cleanest and simplest way to analyze these tricky issues. In this blurb I want to take an alternative  approach and use a simple example to argue:

Claim. Taking into consideration the mere fact of your existence is coherent, and at least in certain situations clearly necessary.

Life and death experiment 1. 

Suppose when you were sleeping the evil Dr. Killian performed a bizarre experiment, he flipped a coin, and if it fell heads did nothing but if it fell tails killed you in your sleep. Suppose now you wake up and find out about the experiment, what will you conclude about the coin? Obviously it fell heads, otherwise you would not be alive now. This is a simple case where you correctly used the fact of your (current) existence to come to the right conclusion.

Notice, by the way, that there is absolutely nothing special about you being alive that makes this datum different from any other piece of information you might possess. For example, suppose instead of killing you in your sleep the doctor simply drew a skull and bones on your forehead if the coin fell tails. Your reasoning would work exactly the same: as soon as you realize, by looking in the mirror, that nothing happened to you or your beautiful forehead you will conclude that the coin fell heads.

I hope this barebones example illustrates to you that there is nothing incoherent or spooky in using the fact that you exist if that happens to be relevant to the situation. Of course, you might say: sure, but this case is very special, because you obviously existed and remember yourself from the time before you fell asleep. True, but as I hope to show you with future wacky scenarios, that doesn't matter remotely as much as one might think it does.

The main point I wanted to convey, however, is unaffected by these concerns, namely that it is completely coherent to use the datum of your current existence. As for the Abiogenesis example, we can understand it if we make some small tweaks to our Dr. Killian story. For that, let's go to part 2

12 Comments - Go to bottom

  1. Hi Dmitriy,

    I may be willing to bite these bullets points.
    * I don't agree with the ensemble analysis
    * I don't much like the doomsday argument, but I can't see a problem with the logic. Either there is some subtle point that is being missed or I'll bite the bullet. I don't see it as reason to use the specific fact of my existence (as opposed to the existence of some observer) as evidence.
    * I don't think we need vague arbitrary rules about what counts as an observer or me.
    * I don't much like having to reject TER, but I'm willing to bite the bullet.
    * Let's see about the intuitively correct results. I would say it also leads to intuitively incorrect results (as I would count inferring the multiverse from the mere fact of our existence).
    * It remains to be seen that completely irrelevant information affects the answer in the opposite view.

    I would say your framing of the abiogenesis example has a problem in that it seems to assume that the number of galaxies is finite. If it's infinite, as it may be, then I see no reason at all to prefer M over S. There are an infinite number of chances to produce you either way. Updating credence on the density of infinite opportunities seems wrong.

    But even if there are merely finitely lots of galaxies, personally I see no reason to entertain M over S.

    I don't think we can infer anything about the frequency of life from the mere fact of our existence. Sure, we can make plenty of inferences for other reasons. For example, there are presumably many more planets with simpler forms of life than advanced civilisations since the latter seems more improbable. I think we can also infer that simple life is perhaps not extremely rare from the fact that it appeared relatively early in earth's history. But nothing from the mere fact of our existence.

  2. Hi Alex

    (continuing our conversation from

    > this just means that every instance of a class must confirm the hypothesis.

    Not at all. Fungible just means interchangeable, such that whichever one we pick won't make a difference to the thesis. This means that the members of the class must not have properties that reinforce or weaken the thesis. I'm just trying to be precise, so that you won't catch me out with an example like my Joe Biden one. But the arguments we're actually talking about are nothing like this. In the arguments we're talking about, observers/universes really are fungible. We're not making claims about their particular properties. No given observer or universe has properties that could ever confirm or disconfirm the existence of other completely causally disconnected universes.

    Of course they might have properties such as fine tuning which might affect credences, and in that case we should reason from the class of universes with those observed properties. We should not reason (I claim) from the evidence of having observed this particular specific universe.

    1. Hey DM,

      If I read you correctly, then that is exactly what I am saying.

      What does fungibility with respect to the hypothesis mean?

      In the way you have used it thus far, it means that every particular being used can't make a different, or invalidate the thesis. But the only way we could guarantee that no possible particular evidence can violate your general hypothesis, is if your hypothesis is a logically necessary one.

      If your general hypothesis is contingent, then it logically follows that it is possible for a particular holding some property to violate the reasoning of the hypothesis. So if this principle "Any inference that we make knowing only that some observable is some fungible instance of a class should not be blocked when we know that it is a specific instance of the class"
      is true, then your hypothesis about the class is logically necessary; otherwise you couldn't possibly guarantee that no particular evidence will violate it.

      In other words, if the claim: "class x members hold property y" is not logically necessary, it is contingent. And if it is contingent it could be wrong, and if it could be wrong then, by definition, a member of class x could fail to hold the property y. This in turn would invalidate/block the above hypothesis, contrary to your statement. And notice of course that the multiverse hypothesis is not a logically necessary one, and so observations about fine-tuning in our universe could potentially invalidate it.

      Regarding this second point that you brought up "No given observer or universe has properties that could ever confirm or disconfirm the existence of other completely causally disconnected universes"

      This is just a claim about irrelevancy; you're saying that our particular evidence can't possibly have any bearing on the multiverse hypothesis being more or less likely to be true because of causal disconnectedness. First of all, there definitely is causal connectedness; it's just that this causal connectedness goes one way. The multiverse caused our universe and us to exist, but we can't impact the rest of the multiverse causally.

      However, one-way causal connectedness is enough. We can see evidence of stars existing in the past which are currently beyond the cosmological event horizon, and we lack the ability to ever causally influence such matter, but that doesn't stop us from making reasonable inferences about the state of such matter in the past or present. For instance, we can reason 'this star we see that existed 10 billion years ago should be going supernova now', even though it is now beyond the event horizon. In this case, we're making a claim about some now causally disconnected space.

      Further, I disagree that one needs causal connectedness to reason about something. It is totally legitimate to use the evidence of our existence for instance, to argue against multiverse models which output high numbers of Boltzmann brains.

    2. Hi Dmitriy,

      > But I am curious if the principle is disconfirmed by the example with 100 sleeping patients and the doctor killing 99 if he flips tails.

      Yes, that's a tough example all right.

      Objectively, we can see that the fact that some patient has survived should not be evidence either way.

      Following the principle of fungibility, just knowing the name of some patient should also not be evidence either way.

      And yet, from the perspective of a particular patient, I would agree that that patient should believe M. And I admit that this poses an undesirable tension with my view. I'm going to more or less contradict myself now, but I'll reconcile what I'm saying after: the difference is that the patient is very special to him/herself, so the observation has a very special property "being me" that is not satisfied by other members of the class, such that the class of fungible members is just the class of objects satisfying the property of "being me", i.e. just me.

      Of course I said earlier that we shouldn't take just being me as being special. The difference here is whether we're talking about post hoc or ex ante reasoning. In the multiverse scenario, we're engaged in post hoc reasoning, as we didn't exist before the universe came into being. In Dr K, we're engaged in ex ante reasoning, as we can make our predictions before the coin is flipped. Or in case we only learn of the experiment after the fact, we can retrodict what we ought to have predicted.

      As you've pointed out before, you can smoothly transition from a post hoc to an ex ante scenario, which poses a problem for my position. But perhaps you can adjust your credences accordingly. To the extent that this looks like a post hoc situation, we regard our identity as unspecial. To the extend that it looks like an ex ante situation, we can regard our identity as special. In between, perhaps we should have some sort of weighted blend.

      After I proposed my "fungibility" idea I came across this paper which has perhaps a better proposal, the Predesignation Requirement. This perhaps helps to explain what I mean about post hoc and ex ante.

    3. Hi Alex,

      > But the only way we could guarantee that no possible particular evidence can violate your general hypothesis, is if your hypothesis is a logically necessary one.

      No, I think you have the wrong idea.

      Hypothesis: There exist lots of atoms.
      Generic observation: I have observed an atom.
      Specific observation: I have observed a uranium atom named Fred.

      The hypothesis is not logically necessary. The generic observation is fungible because the properties of the particular observation have nothing to do with the thesis. They can even be completely irrelevant.

      Hypothesis: Cold fusion will never be realised.
      Generic observation: I have seen a cat.
      Specific observation: I have seen a cat named Whiskers.

      Once you get into making specific observations that have properties relevant to the hypothesis, then I claim you should narrow your class to the set of objects with those observed properties, and no narrower. If you have narrowed it sufficiently, then the members of the class will be fungible with respect to the thesis.

      I feel like the causal disconnectedness discussion is one tangent too many. I'm just trying to rule out for simplicity scenarios where we indirectly observe another universe, e.g. by ripples in the cosmic microwave background.

    4. I think this talk of fungibility is really unhelpful, all members of a class are fungible in the sense that every member is an instantiation of that class. So it doesn't make sense to speak of fungible objects/classes; rather we should speak of fungible properties. The class of cat for instance has necessary properties which every cat member holds. In this sense, every cat member has "fungible" properties that every other cat member holds (the necessary properties of cathood) and non-fungible ones that only that particular cat holds. So in that sense, I'm not even sure what it could mean that say that an observation, as opposed to a property, is fungible or not fungible.

      You'll have to clarify on that because I don't think either Dmitriy or I understand anything you're saying in that regard.

      It was you who made this claim:

      "Any inference that we make knowing only that some observable is some fungible instance of a class should not be blocked when we know that it is a specific instance of the class"

      And I showed that the only way particular evidence cannot possibly invalidate/block a generic observation about a class property, is if the hypothesis about the class property is logically necessary. If that has nothing to do with what you are saying; then you're not making the claim that a hypothesis about a general entity can never be blocked by an observation of a particular evidence. So what are you saying then?

      If by fungibility, you just mean this:
      "The generic observation is fungible because the properties of the particular observation have nothing to do with the thesis"

      Then it sounds like "fungibility" to you just means the proposition that there is a generic and particular observation side by side, where the generic is fungible provided the particular has no bearing on the thesis the generic observation is being used to prove. If that's what fungibility means (and I have no idea how that relates to the conventional use of the word), then it would seem that your definition just begs the question.

      You're assuming that the particular observation of our universe is irrelevant on account of the generic observation being fungible, but being fungible would apparently just mean that the particular observation is irrelevant (according to your quote); which would just be begging the question.

      Some clarification on your part would be extremely helpful here. What does fungibility mean? What is being interchangeable with respect to what? If generic observations are fungible (because they are interchangeable with the members of their class), then particular ones must be fungible too; because particulars belong to a class as well. Unless you mean that particular observations about non-necessary properties are not fungible.

      But then it's not clear why observations about particulars holding non-fungible properties relating to their class can't overturn probabilistic reasoning based on an observation of a generic class member. In my four-way split radioactive decay example, we saw that the particular observation (this is a plutonium atom) invalidated the probabilistic reasoning based on some atom, even though the property of the atom of 'belonging to the elemental class plutonium' is not a necessary properties of all atoms, and therefore a non-fungible property among the class of atoms. But perhaps you are using the word 'fungible' in an entirely different way.

  3. Hey DM,
    "* I don't think we need vague arbitrary rules about what counts as an observer or me."
    But then you agree that it is in fact a problem:
    "As you've pointed out before, you can smoothly transition from a post hoc to an ex ante scenario, which poses a problem for my position. ..."
    That's pretty much exactly what I meant.

    "the difference is that the patient is very special to him/herself, so the observation has a very special property "being me" that is not satisfied by other members of the class, such that the class of fungible members is just the class of objects satisfying the property of "being me", i.e. just me."
    But then how is this different from using your own existence as evidence? If you are willing to narrow down your class of fungible members (which is the same thing, as far as I can tell, as what is often called the "reference class") to just you, then that's no different from my and Alex's view it seems.

    1. Hi guys,

      I'm not going to to try to explain what I mean by fungibility, even though it seems right to me, because the argument in the paper I mentioned is clearer.

      I'll just say the fungibility is about picking the right reference class. You should pick a reference class as broad as you can while it wouldn't matter for the argument what instance of the class is taken as evidence. If that's confusing, forget it and read the paper, which discusses the idea of picking an appropriate reference class without using the term "fungible".

      The paper is excellent and I recognise in it a very clear articulation of the thoughts I've been struggling to express.

      That link again is

      There's a particular thought experiment in there which is extremely close to the Fred problem but which I think actually works. In the thought experiment, we suppose you've got a net which is guaranteed to catch one of the fish in a lake. Some fish are big, and some are small. We observe that we catch a big fish, which we decide to call Asha, and then we must decide if this observation should increase our credence that there are lots of big fish or if there aren't. The correct conclusion is that we should. But the author argues that reasoning from the fact that we caught Asha specifically would block this inference. So TER (called RTE in this paper) is wrong.

      We can draw parallels to Fred as follows.

      Fred = Asha
      Oracle = Net
      Uranium atom = Fish

      So far so good. But now the parallels start to break down

      Atom decays = Fish is caught/Fish is big?

      So the difference is that in Fred, we will learn the identities of any atoms that happen to decay in a certain time period, and we don't know in advance how many. In Fish, we are going to sample exactly one from the population.

      This difference is enough to make Dmitriy's analysis inapplicable.

      In particular, Dmitriy's analysis said:

      > In other words, if one atom decayed (D1), the chance that its id number would be F is independent of how many other atoms are in the sample.

      This is not true in the Fish scenario, where the equivalent might be something like this

      > if a large fish is caught (D1), the chance that it would be Asha is independent of how many other (big?) fish are in the lake

      This is not true because if there are lots more fish in the lake, the chance of catching Asha specifically is reduced.

      Another difference between the two thought experiments is that we tend to imagine a population of possible atoms that may or may not be in the sample for Fred, whereas we tend not to think of a population of possible fish that may or may not be in the lake. In the conclusion of the paper the author actually addresses something like the point I used to explain your Fred analysis to myself -- that a larger sample would be more likely to have Fred in it, and explains why this point isn't so easily applied to the Fish scenario.

  4. Hey DM, I will read that paper, but I wonder whether your response to the Dr.K scenario puts us in agreement, per my previous comment.

    1. To clarify, you are not disputing that it is coherent and at least sometimes necessary to use the evidence of your existence (limit the reference class to just you, as you are willing to do in the Dr. K case). Let me know if that's true.

      But you are disagreeing that we should always do that. Your current position is, if I understand it correctly:
      - if I existed before the experiment, I can consider myself special and put only myself in the reference class (your ex ante case)
      - if I didn't preexist, I should put myself in a wider reference class (post hoc case)
      - there is a smooth, not sharp, transition between those cases

      Is this a fair summary?

    2. I agree with most of that. The paper discusses it in greater clarity than I can.

      I'm less sure that there is a smooth transition between the two cases. If you can show that there is, I'm willing to bite the bullet. But I'm also open to there being some more objective criterion I'm not seeing.

  5. Hey DM, I read the paper, I think the fish scenario doesn't work either, I put my explanation in a new blog post:


Post a Comment