On the Probability of Extraterrestrial Intelligence

Introduction

Speculation on the possible existence of intelligent life elsewhere in the universe appears to have increased in recent years, possibly fueled by several scientific and/or cultural developments, including reports of microscopic life on Mars, popular science-fiction movies (such as Contact and Independence Day), and even initiatives by the Vatican aimed at spreading Christianity to the rest of the cosmos.

Countless internet sites now exist related to this issue, some even warning of the danger of communicating with possibly hostile alien life forms.  Thousands of personal computer owners have donated their spare processor cycles for running the popular SETI@home software to aid in the search for alien signals in massive volumes of radiotelescope data [SETI@home 1999].

Many people are clearly fascinated by the possibility that we may one day discover and possibly even communicate with alien species.  The very existence of extraterrestrial intelligence would have tremendous implications for science, especially for biology.  It is a very exciting possibility.

This excitement, however, must be tempered by a healthy dose of skepticism toward the popular attitude on this issue.  Several arguments in particular have seen repeated use in discussion on this topic, yet it is not difficult to uncover the naivety of some of these lines of reasoning by appealing to some basic mathematical principles.  It is the purpose of this essay to do just that.
 

1. Our existence proves that intelligent life is not improbable.

Perhaps the most oft-cited fallacy is the notion that our very existence serves as a testament to the likelihood of intelligent life existing on any given planet, or in any given solar system having a suitable planet.  To the uninitiated, the statistical nature of this claim has the sound of mathematical sophistication.  It's also intuitively appealing.  But it's wrong.

The first problem with this statement is a minor technical issue involving the use of the word proves.  In purely mathematical domains involving axiomatic systems, it is possible to rigorously prove theorems within those systems, but in science, it is customary to avoid use of the word proves and instead use the word suggests.

This is especially true when interpreting the results of a statistical test.  If a statistical test indicates that the observed phenomena should occur very rarely under a given hypothesis, one is justified in rejecting the hypothesis; otherwise we fail to reject the hypothesis.  Thus, in an experimental setting, an hypothesis is either rejected or not rejected; it is not proven [Zar 1996; Ott 1988].  If as a result of an experiment a scientist fails to reject a theory, we say that the results suggest that the theory is correct, especially if the experiment is replicated many times by different researchers, using different data.

The reason an experiment suggests rather than proves or disproves a theory is that when making an argument based on statistical reasoning, we have to acknowledge that improbable events are occasionally observed.  When a statistical test indicates that an event should be rare under a given set of assumptions, this suggests that our assumptions were not all correct, but it is also remotely possible that our assumptions were correct and that we simply observed a rare event.  In general, it is not possible to distinguish between these two possibilities, so we cautiously favor the former interpretation, with the understanding that future evidence may require a revision of our interpretation.

This is explained in virtually every statistics book:

It is very important to realize that a true null hypothesis occasionally will be rejected, which of course means that we have committed an error in drawing a conclusion about the sampled population. [Zar 1996, p81]

These minor technicalities are significantly overshadowed, however, by the real problem with the argument, which is that our assessment of probabilities based on our observations of life on this planet are highly biased, because we are a part of that system, and we are able to perform these observations precisely because intelligent life did arise on Earth.

This is best illustrated with an example.

Imagine that you are an agent sent by the "Supreme Being" to estimate the prevalence of intelligent life in the universe.  Suppose that you decide to conduct your survey by taking a random sample of planets or solar systems, counting the number of intelligent societies that you discover, and then estimating the overall frequency of intelligence by dividing your count by the sample size.  This is a valid approach to estimation, as long as your sample size is reasonably large.

Now suppose that as you begin your sampling, you are surprised to discover that the very first planet you sample hosts an intelligent life form.  You would not be too unjustified in interpreting this event as suggestive of a high prevalence of intelligent life in the universe, though it would be good to be cautious, because your sample size at this point is only 1, and such small samples are highly susceptible to sampling error.

Intuitively, however, this is compelling evidence that intelligent life is not rare.  The fallacy takes advantage of this intuition.  What invalidates this reasoning in the present case is that we did not perform random sampling when observing life on this planet.  Our species evolved on this planet, and we are unable at present to live elsewhere, so our selection of Earth as the only point in our 1-planet sample is highly biased.  If we could assay only one planet, that planet should definitely not be Earth, because the results of our assay were predetermined by our very existence.

Inferences made on such highly biased sampling strategies are completely invalid [Zar 1996; Ott 1988]; thus the statement that our planet is evidence of a high prevalence of intelligent life in the universe is a fallacy.  Extraterrestrial intelligence may or may not be common, but the present line of reasoning sheds no light on the matter.
 

2. Intelligent creatures are very complex, and thus improbable.

 Intelligent systems are obviously very complicated.  They seem to require a certain, nonrandom organization of computational components, such as neurons, or perhaps transistors [Gell-Mann 1994].

If we consider all the ways that bits of matter (atoms, for example) could be put together, taking note of those combinations which result in intelligence, it seems reasonable to suggest that the intelligent combinations would make up the vast minority of combinations considered.  That is, out of all the random conglomerations of matter that could occur, only a vanishingly small percentage could operate as intelligent biological organisms.

Therefore, as the argument goes, intelligent organisms are highly unlikely to occur at all.  We are lucky enough to exist ourselves; we should not expect to discover yet other intelligent organisms.

This line of reasoning is highly flawed, and it is easy to see why.  Natural systems are not built by randomly piecing together bits of matter.  In the case of biological organisms, they are designed  by the process of natural selection.  Although evolution by natural selection does entail an element of randomness, it is not a completely random process [Koza 1992].  Indeed, it is dintinctly non-random.

Evolution by natural selection works by taking all of the phenotypes (body-plans) currently in existence and mutating some of them at random, to produce a set of mutants.  If any of these mutants turns out to be better in some important way than the other members of its generation, it will be allowed to replace its inferior competitors; i.e., "survival of the fittest."  As this process is repeated, yet other mutants are produced, some of which are superior to any seen so far.

In this way, evolution can be said to be performing a search for highly-fit phenotypes [Dawkins, 1996, p65; Mitchell 1997].  If we imagine all possible phenotypes being arranged as points in some high-dimensional space, then evolution can be envisioned as a search through this space that proceeds by moving from areas of low fitness toward areas of higher fitness.

The current state of the search is represented by all of the phenotypes currently in existence.  Evolution proceeds in its search from these points to nearby points, where "nearby" is defined in terms of mutation distance (if one phenotype can be obtained from some other phenotype by a single genetic mutation, then we say that these phenotypes are separated by a distance of 1 mutation).  Thus, evolution moves through phenotype space by considering the mutational neighbors of the current generation and using the fittest of these neighbors to decide which subspaces to explore next.

Evolution is therefore conducting a biased search through phenotype space (biased toward areas of higher fitness), not a random search.  Because the search is nonrandom, it is not valid to estimate the probability of finding an intelligent species by simply dividing the number of intelligent phenotypes by the size of the search space, as it would if the search was truly random.

In order to decide whether intelligent creatures are improbable, one would need to consider the dynamics of evolutionary search.  Currently, there is no sufficiently powerful mathematical theory of evolution to allow a precise computation of this probability, so the notion that intelligent creatures are improbable should be viewed as pure speculation, and therefore of little value in settling the matter of extraterrestrial intelligence.
 

3. Evolution is progressive; once life emerges, it will become intelligent.

The final argument that we will consider is based on the premise that evolution is in some sense progressive; that it invariably produces higher levels of organization, and also greater cognitive abilities (for a protracted debate on this issue, see for example [Dawkins 1997] or [Depew 1998]).  The implication is that if life could be shown to be common in the universe, then intelligence would necessarily be common also.  Thus, it is claimed that the question of extraterrestrial intelligence is reduced to the simpler question of extraterrestrial life, of which people generally seem to be more accepting.

In a previous section, we characterized evolution as a form of biased search -- biased toward higher fitness.  The present thesis appeals to this characterization, along with the additional assumptions that (1) intelligence and fitness are roughly equivalent, and (2) given enough time, evolutionary search will maximizes fitness.

It is important to be very clear about what is meant by fitness.  Fitness is a measure of the ability of an organism to pass on its genes; i.e., to produce viable offspring that will themselves survive and reproduce, and so on, ad infinitum [Koza 1992; Holland 1992].  It is best thought of as the number of copies of an organism's genes that are ultimately made.

The reason that this definition of fitness is useful for understanding evolution is because it is the endless competition between genes for replication that gives rise to natural selection, and therefore evolution [Dawkins 1976].  It is important to keep in mind that the entire system is ultimately driven by this war between the replicators, because it illustrates the fact that evolution does not favor intelligence per se; evolution will favor intelligence only inasmuch as it promotes greater replication of the genes that code for  intelligence.

There are two obvious components to fitness: the ability to acquire resources (to fuel all of the activities necessary for reproduction), and the ability to avoid death.  Intuitively, one would expect that a more intelligent creature would be better able to perform these tasks.  If intelligence is defined as the ability of an organism to alter its behavior in ways that improve its fitness, then it should be favored by natural selection.

However, this is not the usual connotation of intelligence when discussing extraterrestrials.  Interest seems to focus primarily on the possibility of discovering a species that possesses not only cognitive ability, but also social sophistication.  We fancy conversing with an alien species about their culture, their history, their religion, etc.  It is less clear that natural selection universally favors the adaptations necessary for a species to be culturally sophisticated.  On Earth, many species of animals possess the capacity for nontrivial culture, yet none save humans have a true spoken language with complex semantics [Deacon 1997].

The other assumption, namely that natural selection maximizes fitness, is more easily addressed, due to recent advances in the field of genetic algorithms.  To pose the problem in terms related to mathematical optimization, we introduce the notion of a fitness landscape [Koza 1992; Kaffman 1993].

If we can imagine somehow compressing the high-dimensional phenotype space described earlier so that it fits on the xy-plane, and if we assume there is a function mapping each point on the plane to a real value representing fitness, then it is not too difficult to imagine this function forming a surface in 3-dimensional space.  This surface is called the fitness landscape, because it is often described in terms of plains, hills, and valleys.  Technically, the entire construction should take place in a very high-dimensional space, but it is easier to imagine it in three dimensions.

If phenotypes differing by only a few mutations are close together in phenotype space, then their corresponding fitness values will generally be close together on the fitness landscape.  Thus, evolution is modelled as moving along from point to point in phenotype space while simultaneously visiting corresponding points on the fitness landscape.  The idea of natural selection being biased toward higher fitness translates into the gradual movement of evolution from low areas on the fitness landscape toward higher areas.  This is called hill climbing, because the search proceeds only along smooth, rising surfaces [Mitchell 1997].

However, because natural selection has no foresight, the search that it performs is local in nature, so that if there are many low hills on the fitness landscape, and relatively fewer high mountains (i.e., if the landscape is rugged), the process can easily get trapped at a local maximum, rather than finding the global peak [Mitchell 1997].  The problem is that the gradual nature of evolution prevents it from leaping over obstacles in the fitness landscape.  Occasional anomalies might enable evolution to leave the local basin of attraction, but there is no guarantee that it won't simply end up at another local maximum.  This sub-optimum wandering may continue indefinitely.

In this way, it can be seen that the ability of evolution to maximize fitness depends very strongly on the particular shape of the fitness landscape, and on the way that evolution  traverses that landscape.

At this point, one might argue that the existence of intelligence on Earth establishes that the probability of discovering an intelligent species during evolutionary search is nonzero, and therefore that given enough space-time, the probability of such a discovery can become arbitrarily high.

However, there are two problems with this reasoning.  The first involves the chemistry on which a biotic system is based.  On Earth, that chemistry involves the molecule Deoxyribonucleic acid, but on other planets, it is conceivable that the first successful replicator to spontaneously emerge and proliferate may be based on a different chemistry, with entirely different properties [Kauffman1993].  In these other systems, the fitness landscape may be quite different, radically changing the probability of particular evolutionary paths.  Indeed, it is even conceivable that these other chemistries might support fewer levels of organization above the replicator, so that intelligence might be totally impossible.

The other issue to consider is that the fitness landscape actually changes over evolutionary time, because organisms adapt not only to their abiotic environment, but also to the other organisms present in the biome.  This is called the Red Queen effect [Kauffman 1993].  Evolution on the surface of a planet consists of a vast co-evolutionary process of continual adaptation to changing fitness landscapes.  Maximal fitness is a moving target.   Intelligence may lie on the path to that maximum at some times and not at others.  The situation is very complex.  Assessing probabilities seems virtually impossible.

Finally, consider that even knowing that the probability of intelligent organisms is nonzero (as it must for us to exist) does not in itself indicate that intelligence is common in the universe, because planets that are hospitable to life can remain so for only a finite amount of time.  If we knew the distribution over these durations, and the distribution over chemistries capable of expressing replicators, and the distribution over replicators capable of supporting high levels of organization, we would still lack the necessary understanding of evolutionary dynamics to accurately assess the overall probability of intelligent life in a given solar system.

Without all of this information, the assertion that the evolution of intelligence is inevitable is entirely empty.
 

Conclusion

Seeing through these fallacies requires an understanding of basic statistical principles, as well as a few more advanced concepts from other fields, including biology, computer science, combinatorics, and machine learning.

The question of whether intelligent life is common in the universe remains unanswered.  It is possible that intelligent life is indeed ubiquitous.  However, naive arguments based on faulty reasoning will not settle the issue.  Furthermore, the data and mathematical tools necessary to properly estimate the probability of extraterrestrial intelligence is not yet available, and may never be.
 

References and Further Reading

Dawkins, Richard.  The Blind Watchmaker.  New York:  W.W. Norton & Company, 1996.

Dawkins, Richards. Human Chauvinism.  Evolution 51(3) pp. 1015-1020, 1997.

Deacon, Terrence.  The Symbolic Species:  The Co-evolution of Language and the Brain.  New York:  W.W. Nortan & Company, 1997.

Depew, David.  Are Evolutionists Still Hooked on Progress? Evolution 52(3) pp. 921-924, 1998.

Gell-Mann, Murray.  The Quark and the Jaguar: Adventures in the Simple and the Complex.  New York:  W.H. Freeman and Company, 1994.

Holland, John.  Adaptation in Natural and Artificial Systems.  Ann Arbor: University of Michigan Press, 1975.

Kauffman, Stuart.  The Origins of Order:  Self-Replication and Selection in Evolution.  New York:  Oxford University Press, 1993.

Koza, John.  Genetic Programming:  On the Programming of Computers by Means of Natural Selection.  Cambridge, MA: The MIT Press, 1992.

Mitchell, Tom M.  Machine Learning.  New York: McGraw-Hill, 1997.

Ott, Lyman.  An Introduction to Statistical Methods and Data Analysis, 3rd ed.  Boston: PWS-KENT Publishing Company, 1988.

SETI@home.  SETI@home: Search for Extraterrestrial Intelligence at Home.  http://setiathome.ssl.berkeley.edu/, 1999

Zar, Jerrold H.  Biostatistical Analysis, 3rd ed.  New Jersey: Prentice Hall, 1996.