It was first proposed by the astrophysicist Brandon Carter in 1983, from which it is sometimes called the Carter catastrophe; the argument was subsequently championed by the philosopher John A. Leslie and has since been independently discovered by J. Richard Gott and Holger Bech Nielsen. Similar principles of eschatology were proposed earlier by Heinz von Foerster, among others. A more general form was given earlier in the Lindy effect, in which for certain phenomena the future life expectancy is proportional to (though not necessarily equal to) the current age, and is based on decreasing mortality rate over time: old things endure.
Denoting by N the total number of humans who were ever or will ever be born, the Copernican principle suggests that any one human is equally likely (along with the other N − 1 humans) to find themselves at any position n of the total population N, so humans assume that our fractional position f = n/N is uniformly distributed on the interval [0, 1] prior to learning our absolute position.
f is uniformly distributed on (0, 1) even after learning of the absolute position n. That is, for example, there is a 95% chance that f is in the interval (0.05, 1), that is f > 0.05. In other words, we could assume that we could be 95% certain that we would be within the last 95% of all the humans ever to be born. If we know our absolute position n, this argument implies a 95% confident upper bound for N obtained by rearranging n/N > 0.05 to give N < 20n.
If Leslie's figure is used, then 60 billion humans have been born so far, so it can be estimated that there is a 95% chance that the total number of humans N will be less than 20 × 60 billion = 1.2 trillion. Assuming that the world population stabilizes at 10 billion and a life expectancy of 80 years, it can be estimated that the remaining 1140 billion humans will be born in 9120 years. Depending on the projection of world population in the forthcoming centuries, estimates may vary, but the main point of the argument is that it is unlikely that more than 1.2 trillion humans will ever live.
Assume for simplicity that the total number of humans who will ever be born is 60 billion (N1), or 6,000 billion (N2). If there is no prior knowledge of the position that a currently living individual, X, has in the history of humanity, we may instead compute how many humans were born before X, and arrive at (say) 59,854,795,447, which would roughly place X amongst the first 60 billion humans who have ever lived.
It is possible to sum the probabilities for each value of N and therefore to compute a statistical 'confidence limit' on N. For example, taking the numbers above, it is 99% certain that N is smaller than 6,000 billion.
Note that as remarked above, this argument assumes that the prior probability for N is flat, or 50% for N1 and 50% for N2 in the absence of any information about X. On the other hand, it is possible to conclude, given X, that N2 is more likely than N1, if a different prior is used for N. More precisely, Bayes' theorem tells us that P(N|X)=P(X|N)P(N)/P(X), and the conservative application of the Copernican principle tells us only how to calculate P(X|N). Taking P(X) to be flat, we still have to make an assumption about the prior probability P(N) that the total number of humans is N. If we conclude that N2 is much more likely than N1 (for example, because producing a larger population takes more time, increasing the chance that a low-probability but cataclysmic natural event will take place in that time), then P(X|N) can become more heavily weighted towards the bigger value of N. A further, more detailed discussion, as well as relevant distributions P(N), are given below in the Rebuttals section.
The Doomsday argument does not say that humanity cannot or will not exist indefinitely. It does not put any upper limit on the number of humans that will ever exist, nor provide a date for when humanity will become extinct. An abbreviated form of the argument does make these claims, by confusing probability with certainty. However, the actual conclusion for the version used above is that there is a 95% chance of extinction within 9,120 years, and a 5% chance that some humans will still be alive at the end of that period. (The precise numbers vary among specific Doomsday arguments.)
This argument has generated a lively philosophical debate, and no consensus has yet emerged on its solution. The variants described below produce the DA by separate derivations.
Since Gott specifies the prior distribution of total humans, P(N), Bayes's theorem and the principle of indifference alone give us P(N|n), the probability of N humans being born if n is a random draw from N:
The unconditioned n distribution of the current population is identical to the vague prior N probability density function, so:
giving P (N | n) for each specific N (through a substitution into the posterior probability equation):
The easiest way to produce the doomsday estimate with a given confidence (say 95%) is to pretend that N is a continuous variable (since it is very large) and integrate over the probability density from N = n to N = Z. (This will give a function for the probability that N ≤ Z):
Defining Z = 20n gives:
This is the simplest Bayesian derivation of the Doomsday Argument:
The use of a vague prior distribution seems well-motivated as it assumes as little knowledge as possible about N, given that any particular function must be chosen. It is equivalent to the assumption that the probability density of one's fractional position remains uniformly distributed even after learning of one's absolute position (n).
Gott's 'reference class' in his original 1993 paper was not the number of births, but the number of years 'humans' had existed as a species, which he put at 200,000. Also, Gott tried to give a 95% confidence interval between a minimum survival time and a maximum. Because of the 2.5% chance that he gives to underestimating the minimum, he has only a 2.5% chance of overestimating the maximum. This equates to 97.5% confidence that extinction occurs before the upper boundary of his confidence interval, which can be used in the integral above with Z = 40n, and n = 200,000 years:
This is how Gott produces a 97.5% confidence of extinction within N ≤ 8,000,000 years. The number he quoted was the likely time remaining, N − n = 7.8 million years. This was much higher than the temporal confidence bound produced by counting births, because it applied the principle of indifference to time. (Producing different estimates by sampling different parameters in the same hypothesis is Bertrand's paradox.) Similarly, there is a 97.5% chance that the present lies in the first 97.5% of human history, so there is a 97.5% chance that the total lifespan of humanity will be at least
in other words, Gott's argument gives a 95% confidence that humans will go extinct between 5,100 and 7.8 million years in the future.
Leslie's argument differs from Gott's version in that he does not assume a vague prior probability distribution for N. Instead he argues that the force of the Doomsday Argument resides purely in the increased probability of an early Doomsday once you take into account your birth position, regardless of your prior probability distribution for N. He calls this the probability shift.
Heinz von Foerster argued that humanity's abilities to construct societies, civilizations and technologies do not result in self inhibition. Rather, societies' success varies directly with population size. Von Foerster found that this model fit some 25 data points from the birth of Jesus to 1958, with only 7% of the variance left unexplained. Several follow-up letters (1961, 1962, …) were published in Science showing that von Foerster's equation was still on track. The data continued to fit up until 1973. The most remarkable thing about von Foerster's model was it predicted that the human population would reach infinity or a mathematical singularity, on Friday, November 13, 2026. In fact, von Foerster did not imply that the world population on that day could actually become infinite. The real implication was that the world population growth pattern followed for many centuries prior to 1960 was about to come to an end and be transformed into a radically different pattern. Note that this prediction began to be fulfilled just in a few years after the "Doomsday" was published.
One of the major areas of Doomsday Argument debate is the reference class from which n is drawn, and of which N is the ultimate size. The 'standard' Doomsday Argument hypothesis doesn't spend very much time on this point, and simply says that the reference class is the number of 'humans'. Given that you are human, the Copernican principle could be applied to ask if you were born unusually early, but the grouping of 'human' has been widely challenged on practical and philosophical grounds. Nick Bostrom has argued that consciousness is (part of) the discriminator between what is in and what is out of the reference class, and that extraterrestrial intelligences might affect the calculation dramatically.
The following sub-sections relate to different suggested reference classes, each of which has had the standard Doomsday Argument applied to it.
The Doomsday clock shows the expected time to nuclear doomsday by the judgment of an expert board, rather than a Bayesian model. If the twelve hours of the clock symbolize the lifespan of the human species, its current time of 23:58 implies that we are among the last 1% of people who will ever be born (i.e., that n > 0.99N). J. Richard Gott's temporal version of the Doomsday argument (DA) would require very strong prior evidence to overcome the improbability of being born in such a special time.
The scientists' warning can be reconciled with the DA, however. The Doomsday clock specifically estimates the proximity of atomic self-destruction—which has only been possible for about seventy years. If doomsday requires nuclear weaponry then the Doomsday Argument 'reference class' is people contemporaneous with nuclear weapons. In this model, the number of people living through, or born after, Hiroshima is n, and the number of people who ever will is N. Applying Gott's DA to these variable definitions gives a 50% chance of doomsday within 50 years.
If your life is randomly selected from all lives lived under the shadow of the bomb, this simple model gives a 95% chance of doomsday within 1000 years.
The scientists' recent use of moving the clock forward to warn of the dangers posed by global warming muddles this reasoning, however.
Nick Bostrom, considering observation selection effects, has produced a Self-Sampling Assumption (SSA): "that you should think of yourself as if you were a random observer from a suitable reference class". If the 'reference class' is the set of humans to ever be born, this gives N < 20n with 95% confidence (the standard Doomsday argument). However, he has refined this idea to apply to observer-moments rather than just observers. He has formalized this ( as:
An application of the principle underlying SSSA (though this application is nowhere expressly articulated by Bostrom), is: If the minute in which you read this article is randomly selected from every minute in every human's lifespan then (with 95% confidence) this event has occurred after the first 5% of human observer-moments. If the mean lifespan in the future is twice the historic mean lifespan, this implies 95% confidence that N < 10n (the average future human will account for twice the observer-moments of the average historic human). Therefore, the 95th percentile extinction-time estimate in this version is 4560 years.
If one agrees with the statistical methods, still disagreeing with the Doomsday argument (DA) implies that:
Therefore, these rebuttals try to give reasons for believing that the currently living humans are some of the earliest beings.
For instance, if one is a member of 50,000 people in a collaborative project, the Doomsday Argument implies a 95% chance that there will never be more than a million members of that project. This can be refuted if one's other characteristics are typical of the early adopter. The mainstream of potential users will prefer to be involved when the project is nearly complete. If one were to enjoy the project's incompleteness, it is already known that he or she is unusual, prior to the discovery of his or her early involvement.
If one has measurable attributes that set one apart from the typical long run user, the project DA can be refuted based on the fact that one could expect to be within the first 5% of members, a priori. The analogy to the total-human-population form of the argument is: confidence in a prediction of the distribution of human characteristics that places modern and historic humans outside the mainstream implies that it is already known, before examining n, that it is likely to be very early in N.
For example, if one is certain that 99% of humans who will ever live will be cyborgs, but that only a negligible fraction of humans who have been born to date are cyborgs, one could be equally certain that at least one hundred times as many people remain to be born as have been.
Robin Hanson's paper sums up these criticisms of the DA:
All else is not equal; we have good reasons for thinking we are not randomly selected humans from all who will ever live.
The a posteriori observation that extinction level events are rare could be offered as evidence that the DA's predictions are implausible; typically, extinctions of dominant species happen less often than once in a million years. Therefore, it is argued that human extinction is unlikely within the next ten millennia. (Another probabilistic argument, drawing a different conclusion than the DA.)
In Bayesian terms, this response to the DA says that our knowledge of history (or ability to prevent disaster) produces a prior marginal for N with a minimum value in the trillions. If N is distributed uniformly from 1012 to 1013, for example, then the probability of N < 1,200 billion inferred from n = 60 billion will be extremely small. This is an equally impeccable Bayesian calculation, rejecting the Copernican principle on the grounds that we must be 'special observers' since there is no likely mechanism for humanity to go extinct within the next hundred thousand years.
This response is accused of overlooking the technological threats to humanity's survival, to which earlier life was not subject, and is specifically rejected by most of the DA's academic critics (arguably excepting Robin Hanson).
Here, c and q are constants. If q is large, then our 95% confidence upper bound is on the uniform draw, not the exponential value of N.
The best way to compare this with Gott's Bayesian argument is to flatten the distribution from the vague prior by having the probability fall off more slowly with N (than inverse proportionally). This corresponds to the idea that humanity's growth may be exponential in time with doomsday having a vague prior probability density function in time. This would mean than N, the last birth, would have a distribution looking like the following:
This prior N distribution is all that is required (with the principle of indifference) to produce the inference of N from n, and this is done in an identical way to the standard case, as described by Gott (equivalent to = 1 in this distribution):
Substituting into the posterior probability equation):
Integrating the probability of any N above xn:
For example, if x = 20, and = 0.5, this becomes:
Therefore, with this prior, the chance of a trillion births is well over 20%, rather than the 5% chance given by the standard DA. If is reduced further by assuming a flatter prior N distribution, then the limits on N given by n become weaker. An of one reproduces Gott's calculation with a birth reference class, and around 0.5 could approximate his temporal confidence interval calculation (if the population were expanding exponentially). As (gets smaller) n becomes less and less informative about N. In the limit this distribution approaches an (unbounded) uniform distribution, where all values of N are equally likely. This is Page et al.'s "Assumption 3", which they find few reasons to reject, a priori. (Although all distributions with are improper priors, this applies to Gott's vague-prior distribution also, and they can all be converted to produce proper integrals by postulating a finite upper population limit.) Since the probability of reaching a population of size 2N is usually thought of as the chance of reaching N multiplied by the survival probability from N to 2N it seems that Pr(N) must be a monotonically decreasing function of N, but this doesn't necessarily require an inverse proportionality.
For a similar example of counterintuitive infinite expectations, see the St. Petersburg paradox.
One objection is that the possibility of a human existing at all depends on how many humans will ever exist (N). If this is a high number, then the possibility of their existing is higher than if only a few humans will ever exist. Since they do indeed exist, this is evidence that the number of humans that will ever exist is high.
This objection, originally by Dennis Dieks (1992), is now known by Nick Bostrom's name for it: the "Self-Indication Assumption objection". It can be shown that some SIAs prevent any inference of N from n (the current population).
He gives a number of examples to argue that Gott's rule is implausible. For instance, he says, imagine stumbling into a birthday party, about which you know nothing:
Your friendly enquiry about the age of the celebrant elicits the reply that she is celebrating her (tp = ) 50th birthday. According to Gott, you can predict with 95% confidence that the woman will survive between /39 = 1.28 years and 39[×50] = 1,950 years into the future. Since the wide range encompasses reasonable expectations regarding the woman's survival, it might not seem so bad, till one realizes that [Gott's rule] predicts that with probability 1/2 the woman will survive beyond 100 years old and with probability 1/3 beyond 150. Few of us would want to bet on the woman's survival using Gott's rule. (See Caves' online paper below.)
Although this example exposes a weakness in J. Richard Gott's "Copernicus method" DA (that he does not specify when the "Copernicus method" can be applied) it is not precisely analogous with the modern DA; epistemological refinements of Gott's argument by philosophers such as Nick Bostrom specify that:
Careful DA variants specified with this rule aren't shown implausible by Caves' "Old Lady" example above, because, the woman's age is given prior to the estimate of her lifespan. Since human age gives an estimate of survival time (via actuarial tables) Caves' Birthday party age-estimate could not fall into the class of DA problems defined with this proviso.
To produce a comparable "Birthday party example" of the carefully specified Bayesian DA we would need to completely exclude all prior knowledge of likely human life spans; in principle this could be done (e.g.: hypothetical Amnesia chamber). However, this would remove the modified example from everyday experience. To keep it in the everyday realm the lady's age must be hidden prior to the survival estimate being made. (Although this is no longer exactly the DA, it is much more comparable to it.)
Without knowing the lady’s age, the DA reasoning produces a rule to convert the birthday (n) into a maximum lifespan with 50% confidence (N). Gott's Copernicus method rule is simply: Prob (N < 2n) = 50%. How accurate would this estimate turn out to be? Western demographics are now fairly uniform across ages, so a random birthday (n) could be (very roughly) approximated by a U(0,M] draw where M is the maximum lifespan in the census. In this 'flat' model, everyone shares the same lifespan so N = M. If n happens to be less than (M)/2 then Gott's 2n estimate of N will be under M, its true figure. The other half of the time 2n underestimates M, and in this case (the one Caves highlights in his example) the subject will die before the 2n estimate is reached. In this 'flat demographics' model Gott's 50% confidence figure is proven right 50% of the time.
Some philosophers have suggested that only people who have contemplated the Doomsday argument (DA) belong in the reference class 'human'. If that is the appropriate reference class, Carter defied his own prediction when he first described the argument (to the Royal Society). A member present could have argued thus:
Presently, only one person in the world understands the Doomsday argument, so by its own logic there is a 95% chance that it is a minor problem which will only ever interest twenty people, and I should ignore it.
Jeff Dewynne and Professor Peter Landsberg suggested that this line of reasoning will create a paradox for the Doomsday argument:
If a member did pass such a comment, it would indicate that they understood the DA sufficiently well that in fact 2 people could be considered to understand it, and thus there would be a 5% chance that 40 or more people would actually be interested. Also, of course, ignoring something because you only expect a small number of people to be interested in it is extremely short sighted—if this approach were to be taken, nothing new would ever be explored, if we assume no a priori knowledge of the nature of interest and attentional mechanisms.
Additionally, it should be considered that because Carter did present and describe his argument, in which case the people to whom he explained it did contemplate the DA, as it was inevitable, the conclusion could then be drawn that in the moment of explanation Carter created the basis for his own prediction.
Various authors have argued that the doomsday argument rests on an incorrect conflation of future duration with total duration. This occurs in the specification of the two time periods as "doom soon" and "doom deferred" which means that both periods are selected to occur after the observed value of the birth order. A rebuttal in Pisaturo (2009) argues that the Doomsday Argument relies on the equivalent of this equation:
Pisaturo then observes:
Pisaturo takes numerical examples based on two possible corrections to this equation: considering only future durations, and considering only total durations. In both cases, he concludes that the Doomsday Argument’s claim, that there is a ‘Bayesian shift’ in favor of the shorter future duration, is fallacious.
This argument is also echoed in O'Neill (2014). In this work O'Neill argues that a unidirectional "Bayesian Shift" is an impossibility within the standard formulation of probability theory and is contradictory to the rules of probability. As with Pisaturo, he argues that the doomsday argument conflates future duration with total duration by specification of doom times that occur after the observed birth order. According to O'Neill: