The Dream World of William Dembski's Creationism
By Mark
Perakh
Posted August 19, 2005
The inordinately wellfinanced
Center for Science and Culture of the Discovery Institute of Seattle is the
home of the new antievolution gang. They fight for modifying the school
curricula by inserting creationism as an alternative to evolution, or for what
they euphemistically call "teaching the controversy," yet shrug off the label
of creationism, calling themselves instead Intelligent Design (ID) theorists.
Repeated defeats
of creationists by the US legal system has forced them to regroup and look for
new strategies. ID advocates sport scientific degrees from good universities
and often display substantial erudition and seeming sophistication much
exceeding that of earlier creationists. Since ID purports to be a scientific
enterprise, they need flag bearers with seemingly impressive scientific
credentials, if not actual scientific achievements. Foremost among IDers is
William A. Dembski, with a long list of degrees including a Ph.D. in mathematics,
a Ph.D. in philosophy, and a Master's degree in theology. [1]
Dembski's many
degrees and scores of published books and papers cannot conceal, however, that
he has never conducted real scientific research. Moreover, Dembski's literary
production contains no real mathematics but instead a lot of philosophizing,
often saturated with unnecessary mathematical symbolism. As his extensive
literary production is critiqued by experts, Dembski, without admitting errors,
often surreptitiously shifts his position. These tactics may be handy if
winning the battle regardless of means is the only goal, but they also lead to
the inconsistency that has become Dembski's trademark.
In this article
I shall concentrate on the most salient features of Dembski's prolific literary
output, almost all of which turns out to be poorly substantiated,
contradictory, and often selfaggrandizing.
Dembski's Literary Output, Its Admirers and
Detractors
Dembski's
cotravelers frequently and excessively acclaim his works as great
breakthroughs in science, even to the extent of comparing him to Isaac Newton. [2]
His works, however, have also been extensively criticized. [3] Dembski's latest book is The Design Revolution: Answering the Toughest
Questions about Intelligent Design. [4] Contrary to what this title might suggest, we find in
that book's Index none of the names of his harshest critics. Dembski's book
more properly should have been subtitled Dodging the Toughest Questions about Intelligent
Design.
Dembski is
selective in deciding which critique to respond to and which to ignore. For
example, his (mis)use of the No Free Lunch (NFL) theorems (discussed below) was
subjected to a strong critique by Wein [5] and Wolpert. [6] In
two lengthy rebuttals [7] Dembski spared no effort to reply to Wein,
but he never uttered a word in response to Wolpert. It is not hard to
understand why. Wein, as Dembski stresses in his posts, has only a bachelor's
degree in statistics. This irrelevant factoid, according to Dembski, makes Wein
insufficiently qualified to dispute Dembski's work (in fact Dembski evaded
answering the substance of Wein's well justified critical comments). Dembski
could not use such silly arguments against Wolpert, because Wolpert is a highly
respected mathematician and a coauthor of the very NFL theorems Dembski
misuses.
I also have
written at length about Dembski's ideas. [8] Dembski, who is aware of
my critique [9] has never responded to it. In scientific debates, if
one of the disputants fails to respond, it is usually construed as a silent
admission of errors.
If Dembski
responded to my critique, he could say that I unduly simplify his sophisticated
arguments, thus depriving them of their real deeper meaning. Indeed, I
deliberately simplify his arguments because to my mind all their seeming
sophistication is just a smoke screen intended to make his often hackneyed
notions look like important insights where there is none. I see one of my
tasks, when discussing Dembski's work, as removing its veneer of sophistication
and laying bare his real underlying notions.
Explanatory Filter
Dembski's
Explanatory Filter (EF) is, he claims, a reliable tool for discriminating
between those events that are results of "Intelligent Design" and those that
happen either by chance or due to necessity (also called "law" or "regularity"). He has published a description of the EF in four books and a
number of articles, [10] not to
mention frequent posts to the Internet, so this is obviously an important
component to his ID theory. So far, however, there are no reported instances of
a successful application of the EF by anybody, including Dembski's colleagues
or Dembski himself, to any specific problem where design may be suspected as an
event's antecedent. It is not hard to understand why.
According to
Dembski, there are three and only three clearly distinctive categories of
antecedent factors (causes) for any event: necessity (also called law or
regularity), chance, and Intelligent Design. The term "intelligent" in
Dembski's interpretation does not necessarily imply that the designer possesses
a high intelligence; he (she, it?) even can be stupid; [11] the term
"intelligent" implies only that, regardless of its optimality, design is a
product of an "intelligent agent" rather than of chance or necessity.
Many events
result from a combination of more than one cause. [12] The notion of only three separate causes
does not jibe with reality. Consider Dembski's own favorite case of an archer
shooting arrows at a target painted on a wall. According to Dembski, if the
archer hits the bull'seye, it is a result of design (the archer's skill) and only of design. In fact, the archer's
skill (design) assures only a certain velocity of the arrow when it leaves the
bow. To hit the bull'seye, the arrow must also follow the laws of mechanics,
which determine its trajectory from the bow to the target. Hence, if the arrow
hits the bull'seye, it is the result of a combination of design and regularity
(law, necessity). Similarly, because of an occasional gust of wind, chance also
may become a factor. All three causes can act simultaneously. [12]
Dembski's EF
does not account for events that have more than one cause, and that is one of
the reasons EF is an inadequate tool for the design inference. There are,
however, more serious fallacies in the EF.
The EF is a flowchart comprising three "nodes" that correspond to three steps
employed to decide whether an event is due to chance, necessity, or design. In
fact, of the three "nodes," the first and the second cannot be practically
used. According to Dembski, in the first node of the EF, the probability of the
event is to be estimated. If it is found to be "large" (Dembski offers no quantitative
bounds for such a determination) the event is attributed to law (necessity,
regularity); if the probability is not "large," necessity is rejected, and the
event passes to the second node of the EF.
In Dembski's
imaginary procedure, an event is attributed to necessity (law, regularity) because its probability is found to be
"large." Since probability can't be "read off the event," in practice we can
only do it in the reverse order—first
determine that the event is caused by necessity (law, regularity) and therefore conclude that its probability
is large.
Likewise in the
second node of the EF, Dembski's schema requires us to reestimate the event's
probability. If it is found to be "intermediate" (Dembski again provides no
quantitative bounds for such a determination), the event is attributed to
chance. If, though, the probability in question is "small," the event passes to
the EF's third node. However, to estimate the probability now, we need certain
knowledge about the event's causal history, as we did in the first node. Dembski's schema prescribes the unrealistic,
sequence of steps—an event's probability is found to be intermediate (how?); therefore
we attribute it to chance. This schema is impossible to apply, because we can't
read probability "off the event." Since the first and the second nodes of
Dembski's EF make no sense, the EF cannot be used in reallife situations.
In the third
node of the EF, the final choice is to be made between chance and design.
(Again, the possibility of a combination of causes is ignored, as well as the
notsorare situations wherein no attribution is feasible because of the
paucity of information.) Dembski's prescription here differs from the two
preceding nodes. First, the probability of the event is estimated assuming it happened by chance (the
procedure here is opposite to that suggested for the first and the second
nodes). If this probability is not very "small," the event is attributed to
chance.
Dembski has
suggested a Universal Probability Bound (UPB) [13] which he chose to be
UPB=1/2×10^{150}. This threshold translates into about 500 bits of information (discussed later). If the
probability of a specified event is over the UPB, that is, if the information
associated with this event is over 500 bits, the event, according to Dembski,
cannot be attributed to chance.
I will not discuss here Dembski's reasoning
for his choice of the particular value of UPB (although this reasoning has a
number of dubious elements) because this value per se is not of a
principal significance for this discourse.
Regarding how
small the probability has to be to qualify as a condition for a design
inference, Dembski is inconsistent. On the one hand, he seems to prescribe
using his UPB as the threshold of a sufficiently small probability. On the
other hand, in many of his examples, he views the probability as small enough
for inferring design even when it is by orders of magnitude larger than the
UPB; for example, he considers a sevendigit phone number as sufficiently
improbable to justify a design inference although the probability of this
number is immensely larger than the UPB. [14]
According to
Dembski, if the probability of a chance occurrence of an event is found to be small,
then the event must next be tested for specification.
Let us construe specification in its most common sense—as a choice of an
object out of a set of similar objects. For example, if you are asked to pull
"a card" from a deck, the card is not "specified." Any card you randomly choose
will do. If, though, you are asked to pull the seven of spades from a deck,
this time the card is specified, and only the seven of spades will do. This
simple example shows the commonly understood difference between specified and
unspecified objects.
What is the
probability that whichever card you randomly choose will turn out to be "a
card"? Obviously this probability is p=1 (or 100%), because any card you choose
meets the definition of being "a card." What is the probability that the card
you randomly pull from a deck will turn out to be the seven of spades? Since
there are 52 cards in the deck, each having the same chance of being randomly
chosen, the probability in question this time is 1/52. Hence, in this example specification
makes the event's probability 52 times lower than for an unspecified event.
In general,
specification always decreases the probability of an event. There is no basis
for construing specification as a category qualitatively independent of the
event's probability (as Dembski suggests). Specification (if any is detected)
adds only quantitatively to the probability estimate but warrants no
qualitatively different contribution to the putative design inference.
Dembski imposes
certain restrictions on specification if it is to be used for a design
inference [15]. Therefore Dembski's specification is a narrower concept
than the one used in the discussion above. Not all events that are specified in
the above broad sense are specified in Dembski's sense. However, all events
that are specified in Dembski's sense all are also specified in the above
broader sense. Therefore, applying the concept of specification only in
Dembski's narrower sense does not invalidate the assertion that specification
quantitatively decreases the event's probability rather than adds a
qualitatively different factor to the design inference.
Therefore the
procedure suggested by Dembski for the third node of his EF boils down to the
estimate of the event's probability, either directly or disguised as
specification. The design inference is
thus reduced to an argument from improbability.
Furthermore,
probability is a quantity whose estimate, as Dembski himself asserts,^{16}
is determined by the knowledge about the event in question (in Dembski's
parlance, by the "background information H"). Obviously the other side of the
coin is the assertion that the estimated probability reflects the level of our
ignorance about the event. Dembski's discourse, including his EF, is just a
feebly disguised argument from ignorance;
its other name is the Godofthe gaps
argument, which has lost credibility even among Dembski's philosophical
colleagues. [17]
Dembski admits
that EF can produce false negatives, that is, fail to detect ID where ID is in
fact present. He insists, however, that his EF does not produce false
positives, that is, if it detects design, this result is beyond doubt. To prove
this assertion, Dembski offers two lines of proof.
1. Dembski's
first proof of EF's reliability in regard to false positives is a "...straightforward inductive argument: in
every instance where the explanatory filter attributes design and where the
underlying causal history is known, it turns out design is present; therefore
design is actually present whenever the explanatory filter attributes design." [18]
First, as
philosopher Dembski must know, if A entails B, B does not necessarily entail A.
Even if his assertion (that the explanatory filter correctly infers design
whenever an event is known to be caused by ID) were true, that fact in itself
would not necessarily lead to the reverse conclusion (that each time
explanatory filter attributes design, intelligence is indeed the causal
antecedent of the observed event). In fact, though, Dembski does not
substantiate even the underlying statement (for example, by providing a more or
less extensive record of those situations where the causal history is known and
filter infers design). He discusses just a few examples, and we cannot know
whether or not these examples were selected at random or chosen deliberately
(cherrypicked) because they seemed to fit his thesis. Moreover, this alleged
proof can be shown to be false by simply pointing to cases where EF clearly
produces false positives. There are many examples of such false positives [19]
and we will discuss one more (the case of triangular snowflakes) below.An early example of false positives came from the ID
camp itself, from a prominent proponent of ID, philosopher of science Del
Ratzsch. [20] Dembski never addressed Ratzsch's example; neither did he
acknowledge the examples of false positives suggested by other critics of his
work, while continuing to insist that his EF never produces false positives. [21]
2. Dembski
second "proof" of the EF's reliability is: "The Explanatory Filter is a
reliable criterion for detecting design because it coincides with how we
recognize intelligent causation generally." [22] If this is so, why do
we need EF as we can detect design without it, and know how to do it, and do it
"generally"? On the other hand, if his EF is indeed a novel tool for detecting
design, which is superior to "how we
do it generally," how can its reliability be asserted by comparing it to an inferior procedure used without it?
Both "proofs" of
EF's reliability suggested by Dembski are in fact not convincing.
Specified Complexity
Dembski pays much attention to
specification as such, apart from its role in the EF, and he employs a number
of alternative terms, such as Complex Specified Information (CSI), Specified
Complexity (SC), and sometimes simply specification, often shifting the meaning
he attributes to these terms.
When using the
term specification, Dembski explains
that it means a certain type of pattern. [23] To serve as a specification,
says Dembski, the pattern must meet a set of additional requirements. The most
important among those requirements is perhaps "detachability." For example, if
you see a heap of stones arranged in a certain pattern that is not familiar to
you, this pattern is not "detachable" from this specific heap of stones—it does
not match any image you have antecedently stored in your memory. If, though, an
astronomer comes across the same heap of stones, and recognizes in it a pattern
reproducing the shape of a certain constellation, the image of which is
familiar to him, then for him this pattern is "detachable" from the particular
heap of stones and serves as a specification (that is, can lead to the design
inference: the conclusion that some intelligent agent has intentionally placed
the stones in the image of a constellation). [24] This example,
however, shows that Dembski's specification is no more than a subjective recognition of the pattern.
(In this example, as well as in many other instances discussed by Dembski, the
probability that the pile of stones has the shape of a constellation may be
many orders of magnitude larger than UPB=1/2 ×10^{150} ; this is just
one of the many examples of Dembski's inconsistency.) Another necessary
component of specification, according to Dembski, is complexity, which he construes as tantamount to low probability:
"Probability measures are disguised complexity measures...with the disguise
involving nothing more than a change in direction and scale." Similarly, "the
greater the complexity, the smaller the probability" [25]
I believe that
the very concept of complexity as disguised improbability is contrary to facts
and logic. For example, under certain (rare) weather conditions, an unusual
triangular shape of snowflakes can be observed. [26] Unlike more common forms of snowflakes with
their intricately complex structure, these rare snowflakes have a simple
structure. As Dembski asserted, [27]
snow crystals' shapes are due to necessity—the laws of physics
predetermine their appearance. However, triangular snowflakes, while indeed
predetermined by laws of physics, occur only under certain weather conditions,
which are very rare and unpredictable. Therefore we have to conclude that the
emergence of the triangular snowflakes is a random event. This is another
example where at least two causal antecedents—chance and law—are in play
simultaneously.
Since the
appropriate weather conditions occur very rarely, the probability of the chance
emergence of the triangular snowflakes is very small; also, they have a
uniquely specific shape. Hence,
according to the EF, these snowflakes were deliberately designed. The more
reasonable conclusion, however, is that they appeared by chance (plus the
necessary contribution of law). (This is also another example of a false
positive produced by the EF.) Since the probability of the occurrence of these
snowflakes is small, then, according to Dembski's insistence that large
complexity is equivalent to low probability, their complexity must be large. In
fact, though, the rare triangular snowflakes have the simplest form among all
the snowflakes observed. Thus, Dembski's thesis asserting that complexity is
tantamount to small probability is an unsubstantiated and therefore misleading
suggestion.
Complex Specified Information
Dembski's Complex Specified Information (CSI) is in
fact a combination of low probability (which he construes as tantamount to
complexity) with a recognizable pattern. For example, according to Dembski, a
string of gibberish displays no CSI, even if its spontaneous emergence has a
very low probability, but a segment of a meaningful text possesses CSI because not
only is its spontaneous emergence improbable but it also displays a
recognizable pattern.
In its formal
rendition, Dembski's Complex Specified Information (CSI) comprises three
components, (a) information, (b) complexity and (c) specification.
(a) Information. Dembski's definition of
information is [28]
I(E)= log_{2} p(E)
...........(1)
where I stands for information associated with an individual event E, p is the probability of that event,
and the logarithm is to the base of 2. In information theory, I is often called surprisal; another, more recent term is selfinformation. [29]
The definition
(1) simply expresses probability in a logarithmic form. In this rendition, the
concept of information contains
nothing beyond the concept of probability.
If a definition
has been selected, it has to be applied consistently. However, having chosen
(1) as a definition of information, a
few pages further [30] Dembski
refers to the same quantity I as
complexity. If I is complexity, then (1) contradicts
Dembski's own earlier definition of complexity [31] as a measure of a
difficulty of solving a problem, since (1) has nothing to do with the
difficulty of a problem.
(b) Complexity, as we have seen, in
Dembski's view is the equivalent of low
probability.
(c) Specification likewise cannot be viewed
as a category qualitatively
distinctive from probability (as discussed earlier).
Hence, all three
components of CSI are in fact just components of the overall probability of the
event whose causal history is in question. Dembski's use of the concept of CSI,
and with it his "complexityspecification criterion," [32] are just a
rephrased argument from improbability,
or, therefore, as noted above, an argument
from ignorance, to which Dembski has added no novel features besides
unnecessary mathematical symbolism.
The Law of Conservation of Information
Dembski's penchant for idiosyncratic
terms and allegedly revolutionary novel concepts perhaps finds its most salient
expression in his Law of Conservation of Information (LCI). Experts in
information theory have so far paid no attention to Dembski's supposedly
revolutionary breakthrough, [2] so that there are no references to
Dembski's LCI in any books or papers professionally dealing with information
theory.
The most concise rendition of LCI by Dembski
is perhaps :"Natural causes are incapable
of generating CSI." And Dembski's first corollary: "The CSI in a closed system of natural causes remains constant or
decreases," [33] may serve as an alternative rendition of LCI insofar
as it is relevant to our discussion.
Dembski does not
define "closed system of natural causes." In particular, it is unclear whether
it includes human intelligence, which obviously can generate CSI but usually is
not considered supernatural. Regardless, I will show that LCI, if followed
consistently using Dembski's notions, contradicts the second law of
thermodynamics. (Apparently not satisfied with introducing a new "law"
regarding information, Dembski suggests that his LCI can be expanded to become
the Fourth Law of Thermodynamics. [34])
Let us see whether
Dembski's various statements, if applied consistently, lead to a conclusion
compatible with the second law of thermodynamics. To this end, let us juxtapose several of Dembski's statements
relevant to his LCI.
(a) Dembski uses the concept of entropy H
according to Shannon's information theory, asserting that "the average
information per character in a string is given by entropy H." [35] That is,
Dembski maintains that
average information = entropy.
(b) CSI, by definition, is a combination of three
components: information, specification, and complexity.
From (a) and (b)
follows that the behavior of CSI (which includes information as its necessary
component) must not contradict those laws which determine the behavior of
entropy. Entropy obeys the second law of thermodynamics according to which in a closed system entropy cannot
spontaneously decrease. Since Dembski accepts that "entropy = average
information" (see above), he must conclude that average information associated with a closed system cannot
spontaneously decrease; it can only increase or remain unchanged.
A
thermodynamically closed system, by definition, does not exchange matter and/or
energy with its surrounding. A system closed in the informational sense does
not exchange information with its surrounding. Although matter/energy and
information are not the same, the behavior of both in many respects can be
characterized by the same quantity—entropy. The units used for thermodynamic
entropy differ from informational entropy, but that is due only to convenience;
entropy is essentially a dimensionless quantity [36] whose behavior is
determined by the same laws both for thermodynamic and informational entropy.
Hence, while
Dembski's LCI asserts that CSI cannot increase
in his "closed system of natural causes," his other notions, if combined with
the second law of thermodynamics, assert that average information (a.k.a. entropy)
in a closed system cannot decrease. These two statements are incompatible.
Specification—one
of the components of CSI—is, in Dembski's rendition, a qualitative concept. An
event can either be specified or not; there are no degrees of specification
(see, however, another view [37]). Two other components of
CSI—information and complexity—are quantitative; however complexity, according
to Dembski, is just a property of information when the latter exceeds about 500
bits. Therefore the increase or decrease of CSI necessarily implies the increase
or decrease of its constituent information (because the third component of
CSI—specification, is only a qualitative concept). Hence, if we talk about an
increase or decrease of CSI, then, in accordance with Demsbki's concepts, we
necessarily talk about an increase or decrease of information, and therefore of
entropy (which is just average information – see above).
Whereas the
second law of thermodynamic prohibits entropy's decrease, Dembski's LCI allows
for its decrease but prohibits its increase. I see no way to reconcile Dembki's
LCI with the second law of thermodynamics. This seems to be a sufficient reason
(albeit not the only reason) to assert that Dembski's alleged fourth law of
thermodynamics, which is supposed to be a generalization of his LCI, makes no
sense.
No Free Lunch?
The title (No Free Lunch) of
Dembski's 2002 book refers to certain theorems of optimization theory^{38}
that Dembski asserts make evolution by a Darwinian path impossible. For
example, "The No Free Lunch theorems dash any hope of generating specified
complexity via evolutionary algorithms" (p.196). Similarly, "The No Free Lunch
theorems show that evolutionary algorithms, apart from careful finetuning by a
programmer, are no better than blind search and thus no better than pure
chance" (p. 212), and "The No Free Lunch theorems show that for evolutionary
algorithms to output CSI they had first to receive a prior input of CSI" (p. 223).
There are
several NFL theorems, the most relevant for our discussion being the "first NFL
theorem for search" (NFL1).
At the heart of
the NFL theorems are two concepts: a fitness
function (or its opposite, a cost
function), and search algorithms.
A fitness function (often a synonym of "figure of merit") is a quantitative
characteristic of a system that is related to its functioning or its
usefulness, or is of interest for any other reason. For example, the fitness
function may list the heights of peaks in some mountainous region as a function
of their locations. The physical relief of the mountainous region is an example
of a fitness landscape. Imagine that
we want to find the highest peak in that region. Search algorithms are the
sequences of steps to be taken in the search for the highest peak. The search
may be directed toward a certain target—say, a peak which is 6,000 meters above
the sea level. In this case the search is terminated when the target, the 6,000
meter high peak, has been located. Equally the search may not be directed
toward a preselected target but may be conducted until, say, a preselected
number of peaks have been explored, and then the search is terminated
regardless of how tall the last conquered peak turns out to be.
NFL1 is equally
valid for the first (targeted) and the second (nontargeted) searches. Algorithms are strategies employed for
the search. One strategy may be climbing up peaks one by one, moving from the
mountain region's periphery toward its geographic center, measuring the heights
of the conquered peaks with an altimeter and recording them. Another strategy
may suggest climbing peak after peak, selecting at each step a nearby peak that
looks higher than the one already conquered. These two strategies correspond to
two different search algorithms.
Although NFL1 per se is not related to any performance
measures, it may be convenient to employ certain performance measures which, although not required by the NFL
theorem, may facilitate the judgment about the efficacy of an algorithm. It is
in principle unimportant how the data directly obtained in the search are
mapped onto the performance measure. [38] The performance measure in a
search for a preselected peak (a targeted search), may be, for example, chosen
as the number of steps an algorithm needs to find the target. The algorithm
that finds the target in fewer steps "performs" better. The performance measure
in a nontargeted search which terminates after a preselected number of steps
may be, for example, the maximum height reached after the preselected number
of climbs. The algorithm that ends the search at a taller peak "performs"
better.
NFL1 per se has nothing to do with the choice
of a performance measure. It considers the set of data obtained by the search
algorithms and presented as a table listing the measured values of the fitness
function in temporal order. This table is called a sample. NFL1 relates to samples in probabilistic terms. It says
that the probabilities of obtaining a
certain sample by two different search algorithms are the same if these
probabilities are averaged over all possible fitness landscapes.
The word averaged is crucial. NFL1 asserts that
no algorithm is better than any other algorithm if their results are averaged over all possible fitness functions.
Thus, if a certain algorithm is better on a certain type of fitness landscapes,
it is necessarily worse on some other types of fitness landscapes. Of critical
importance, the NFL1 theorem says nothing about any algorithms' advantages or
shortcomings on specific fitness landscapes, where one algorithm may
drastically outperform other algorithms at the cost of being inept on some
other fitness landscapes.
Now look at how
Dembski renders the gist of the NFL theorems:^{39}
"A generic NFL theorem now takes the following form: It
sets up a performance measure M that characterizes how effectively an
evolutionary algorithm E locates a target T within m steps using information
j."
This description
is wrong in all of its parts. First, NFL1 does not "set up a performance
measure." No such quantity is mentioned in the theorem's proof, which is valid
regardless of any performance measure. Second, NFL1 does not relate to any
targets. It is equally valid for algorithms searching for a target and for such
algorithms that do not search for any preselected target. Third, NFL1 has
nothing to do with the any information j which resides outside the search
space (this point will be discussed below in the section about the
"displacement problem").
Dembski misuses the NFL
theorem when he asserts that evolutionary algorithms cannot outperform blind search:
"since blind search always constitutes a perfectly valid evolutionary
algorithm, this means that the average performance of any evolutionary
algorithm E is no better than blind search" [39] Since blind search is
an extremely slow process, and no other algorithm can do better that blind
search, then, according to Dembski, evolutionary algorithms cannot ensure the
rate of evolution required by evolution theory, which entails random mutations
plus natural selection.
This statement
is misleading. Evolutionary algorithms indeed cannot outperform blind search
but, as Dembski knows, only if their performance is averaged over all possible fitness functions. They can (and do)
immensely outperform blind search on specific fitness landscapes, both in
computer simulations and in the real biosphere. Dembski himself reviews
examples of evolutionary algorithms immensely outperforming blind search (or
random sampling) — like Richard Dawkins' "weasel algorithm," [40] a checkersplaying algorithm [41],
and an antennadesigning algorithm [42]
— but he forgets about them when wrongly claiming that NFL1
prohibits biological evolution.
Dembski admits
that evolutionary algorithms can outperform blind search if they are finetuned
by a programmer. He does not believe, though, that natural genetic algorithms
can be naturally finetuned to climb natural fitness landscapes.^{43}
This disbelief, although it is Dembski's prerogative, is not based on empirical
or logical foundation but only on Dembski's philosophical/religious convictions.
It is possible, though, to show that the fitness landscapes encountered in the
real biosphere can often be finetuned to the available genetic algorithms
based on mutations and natural selection. Let us discuss it (see the following
section).
Displacement Problem
Chapter 4 in Dembski's
No Free Lunch is devoted to NFL1which
allegedly proves the impossibility of Darwinian evolution. Dembski's thesis was
strongly rebuffed by a number of critics, including the coauthor of the NFL
theorems, David Wolpert. [6] Confronted with critique, Dembski, without
acknowledging his error, tried to make it look inconsequential. Since he could
not make chapter 4 in his book disappear, he announced instead [44]
that, contrary to the obviously triumphant appeal to the NFL theorems in his
book, these theorems were not really crucial for his thesis but just a
particular example of what he calls displacement
problem (which, however, was in fact introduced in his book as a
consequence of his interpretation of the NFL theorems).
Here is how Dembski defines the
displacement problem: [45]
"...the problem of
finding a given target has been displaced to the new problem of finding the
information j capable of locating
that target. Our original problem was finding a certain target within phase
space. Our new problem is finding a certain j within the informationresource
space J."
As noted
earlier, the real problem is not necessarily "finding a given target" because
search algorithms may work without being directed toward a "given target," and
indeed, biological evolution is a process where there is no longterm target.
With his
habitual inconsistency, Dembski in some instances writes that evolutionary
algorithms are supposed to be "non teleological," that is, not directed to a
target, and also that biological evolution is a process without a preselected
target. In other instances, however, he states the opposite — that evolution is
after all at least partially targeted. For example, "evolutionary
algorithms are supposed to be capable of solving complex problems without
invoking teleology." [46]But
later, "An evolutionary algorithm is supposed to find
a target within phase space." [46] Since a search for a target is
necessarily teleological, these two statements are contradictory.
Furthermore, all
Dembski's talk about information j
sitting in an "informationresource space J" is irrelevant to real life
problems. The NFL theorems are only valid for "blackbox algorithms." This
means that, before starting the search over the fitness landscape, the algorithm
incorporates no knowledge whatsoever about the landscape. It probes the
landscape one point at a time, gradually acquiring bits of information about
the landscape. The fitness landscape is always a given: the algorithm has no choice of a fitness landscape but only
explores an existing landscape it faces. Therefore the displacement problem is
a phantom.
Imagine an
organism existing in a certain environment. As a simplified example, let's say
it is an animal that feeds on fruit growing on trees but has no treeclimbing
skill. If the animal is too small, it has problems in reaching the fruit and
its survival is uncertain. If it is too tall, it has problems in moving through
the dense jungle and thus reaching more trees. Hence we might expect a certain
optimal size for that animal, not too small and not too large, which provides
the best chance of surviving and thus having progeny. The performance measure
in this case can be, say, the number of descendants left by the animal, or the
duration of the animal's life, or any other quantity that can be used to
characterize the animal's chance for leaving more descendants. If we plot the
dependence of this quantity on the animal's size, we will get a moreorless
bellshaped curve with a peak of fitness at a certain optimal size. This graph
is a simplified, twodimensional model of a fitness landscape (which usually is
multidimensional). This fitness landscape is determined by the environment.
The animal has no choice of a fitness landscape — it is given. Search algorithms
(sequences of events and actions resulting in exploration of the fitness
landscape) that entail random mutations and natural selection fit in well with
this fitness landscape: those mutations which result in the animal's size
approaching the optimal value will be naturally
selected to ensure the maximum fitness, that is, have the best chance of
leaving progeny.
No excursion
into the "informationresource space" imagined by Dembski is required, so his
displacement problem is an abstract invention that does not exist in practice.
Conclusion
Dembski has either authored or
edited at least eight books and numerous articles, essays, and Internet posts.
He does not shy away from reproducing, often verbatim, the same passages time
and time again, apparently striving for having his ideas disseminated as widely
as possible, in every medium he can reach. On the other hand, encountering
criticisms, he sometimes surreptitiously modifies his argument (without ever
admitting error) so as to quietly slide out from the predicament caused by the
criticism. [47]
Because of the
large volume of Dembski's publications, my review has necessarily had to be
cursory. I hope I have nevertheless shown that Dembski's so highly acclaimed
achievements are just a nebulous dream; the real contents of his ideas and
notions are in an inverse relation to the intensity of the praise heaped upon
him by the ID crowd. If Dembski's work is the best the ID advocates have to
show, then the entire ID enterprise is a political movement that wholly lacks
scientific significance.
Comment: After
this paper was submitted for publication, Dembski posted to the internet two
essays allegedly providing "mathematical foundation of Intelligent Design." A
brief critical discussion of these essay can be seen at http://www.talkreason.org/articles/math.cfm
and http://www.talkreason.org/articles/newmath.cfm.
Acknowledgment.
I would like to
thank Matt Young for helpful editorial advices.
Notes
[1] Dembski, William. 2004. "Biographical Sketch."
http://www.designinference.com/biosketch.htm, accessed on December 12, 2004.
[2] Koons, Rob. 1999. Blurb on the dust cover of Dembski's
Intelligent Design (see note 10c).
[3] (a) Edis, Taner.
2002. "Darwin in Mind: Intelligent Design Meets Artificial Intelligence."
www.csicop.org/si/200103/intelligentdesign.html, accessed on June 23.
(b) Eells, Ellery. 1999. "Review
of The Design Inference by William A. Dembski." Philosophical Books
40, No 4.
(c) Elsberry, Wesley R. 1999.
"Review of WA Dembski, The Design Inference," Talk Reason
http://www.talkreason.org/articles/inference.cfm, accessed on August 14,
2003.
(d) Elsberry, Wesley R. and
Jeffrey Shallit, 2003. "Information Theory, Evolutionary Computation, and
Dembski's Complex Specified Information." Talk Reason.
http://www.talkreason.org/articles/eandsdembski.pdf, accessed on April 29, 2004.
(e) Fitelson, Branden,
Christopher Stephens, and Elliott Sober. 1999."How Not to Detect Design—Critical
Notice: William A. Dembski, The Design Inference." Philosophy of
Science 66: 472–88.
(f) GodfreySmith, Peter. 2001.
"Information and the Argument From Design." In R. T. Pennock, ed.,
Intelligent Design Creationism and Its Critics: Philosophical, Theological and
Scientific Perspectives. Cambridge, MA: MIT Press: 575596.
(g) Korthof, Gert. 2000. "On
the Origin of Information by Means of Intelligent Design", in Was Darwin
Wrong?
http://home.planet.nl/~gkorthof/kortho44.htm, accessed on August 1, 2003.
(h)
Pennock, Robert T. 1999. Tower of Babel: The Evidence against the New
Creationism. Cambridge: MIT Press.
(i) Orr, H Allen, 2002, "Review
of No Free Lunch, by William Dembski." Boston Review, 27,
no. 3.
(j) Pigliucci, Massimo.
2001."Design Yes, Intelligent No: A Critique of Intelligent Design Theory and
NeoCreationism." Skeptical Inquirer 25, no. 5: 34–39.
(k) Ratzsch, Del. 2001.Nature,
Design, and Science: The Status of Design in the Natural World. New York:
State University of New York Press: 153168.
(l)
Rosenhouse, Jason. 2002. "Probability, Optimization Theory, and Evolution."
Evolution, v. 56, No 8, 1721.
(m)
Shallit, Jeffrey. 2003. "Review of Dembski's No Free Lunch."
http://www.math.uwaterloo.ca/~shallit/nflr3.txt, accessed on August 7,
2003.
(n) Shallit, Jeffrey and Wesley
R. Elsberry. 2004. "Playing Games With Probability: Dembski's 'Complex Specified
Information." Chap. 9 of M. Young and T. Edis, eds.Why Intelligent Design
Fails: Scientific Critique of the New Creationism. New Brunswick, NJ: Rutgers University Press.
(o) Shanks, Niall. 2004. God,
the Devil, and Darwin. New York: Oxford University Press.
(p) Stenger, Victor J. 2001.
"Intelligent Design—The New Stealth Creationism," Talk Reason,
http://www.talkreason.org/articles/Stealth.pdf, accessed on June 12, 2003.
(q) Tellgren, Erik. 2002. "On
Dembski's Law of Conservation of Information." Talk Reason.
http://www.talkreason.org/articles/dembski_LCI.pdf, accessed on April 28, 2004.
(r) Van Till, Howard J. 2003. "E
coli at the No Free Lunch Room: Bacterial Flagella and Dembski's case for
Intelligent Design."
www.aaas.org/spp/dser/evolution/perspectives/vantillecoli.pdf, accessed on
April 28, 2004.
(s) Wein, Richard. 2000. "Wrongly
Inferred Design." Talk Reason, http://www.talkreason.org/articles/wrongly.cfm, accessed on April 28, 2004.
(t) Wilkins, John S. and Wesley
R. Elsberry. 2001. "The Advantages of Theft over Toil: The Design Inference and
Arguing From Ignorance." Biology and Philosophy, 16: 711724.
(u)Young, Matt. (u1) 2001.
"Intelligent Design Is Neither," paper
presented at the conference Science and Religion: Are They Compatible?
Atlanta, Georgia, November 911, www.mines.edu/~mmyoung/DesnConf.pdf, accessed on April 28, 2004.
(u2) 2002. "How to Evolve
Specified Complexity by Natural Means,"
www.pcts.org/journal/young2002a.html, accessed on April 28, 2004.
(u3)
2004. "Dembski's Explanatory Filter Delivers a False Positive," Panda's Thumb, http://www.pandasthumb.org/ptarchives/000166.html, posted April 22, 2004.
(And others.)
[4] Dembski, William. 2004. The Design Revolution:
Answering the Toughest Questions about Intelligent Design, Downers
Grove: InterVarsity Press.
[5] Wein, Richard. (a) 2002. "Not a Free Lunch but a Box
of Chocolate," Talk Reason,
http://www.talkreason.org/articles/choc_nfl.cfm, accessed on April 28, 2004.
(b) 2004. "The
DesigneroftheGaps Revisited." In Talk Reason,
http://www.talkreason.org/articles/Designer.cfm, accessed on April 28, 2004.
[6] Wolpert, David H. 2003. "Dembski's
Treatment of the NFL Theorems Is Written in Jello," in Talk Reason,
http://www.talkreason.org/articles/jello.cfm, accessed on April 28, 2004.
[7] Dembski. 2002
(a) "Obsessively Criticized but Scarcely Refuted: A Response
to Richard Wein."
http://www.designinference.com/documents/05.02.resp_to_wein.htm.
Accessed on April 29, 2004.
(b) "The Fantasy Life of Richard Wein: A Response to Response."
www.designinference.com/documents/2002.6WeinsFantasy.htm.
[8] Perakh, Mark. (a) 2001."A Consistent Inconsistency." Talk Reason. http://www.talkreason.org/articles/dembski.cfm, accessed on April 28, 2004.
(b) 2002 "A Free Lunch in a
Mousetrap." Talk Reason.
http://www.talkreason.org/articles/dem_nfl.cfm, accessed on April 28, 2004.
(c) 2003 "A Presentation without
Arguments: Dembski Disappoints." Skeptical Inquirer 26, no. 6: 31–34.
(d) 2003. "The No Free Lunch
Theorems and Their Application To Evolutionary Algorithms." In Talk Reason.
http://www.talkreason.org/articles/orr.cfm, accessed on April 28, 2004.
(e) 2004. Unintelligent Design.
Amherst, NY: Prometheus Books (chapter 1).
(f) 2004. "There Is a Free Lunch
after All: William Dembski's Wrong Answers to Irrelevant Questions." Chapter 11
of M. Young and T. Edis, eds., Why Intelligent Design Fails: A Scientific
Critique of the New Creationism, New Brunswick, N. J.: Rutgers University Press.
(g) 2004. "The Design
Revolution? How William Dembski Is Dodging Questions About Intelligent Design."
In Talk Reason.
http://www.talkreason.org/articles/Revolution.cfm, accessed on April28, 2004.
(h) Elsberry, Wesley, and Mark
Perakh, 2004. "How
Intelligent Design advocates turn the sordid lessons from Soviet and Nazi
history upside down." >Talk
Reason.
http://www.talkreason.org/articles/eandp.cfm, accessed on April 28, 2004.
(i) Perakh, Mark, and Matt Young. 2004. "Is Intelligent Design Science?"
Chap. 13 of M. Young and T. Edis, eds., Why Intelligent Design Fails: A
Scientific Critique of the New Creationism. New Brunswick, N. J.: Rutgers
University Press.
[9] Dembski, William. 2004. (No
title).
http://www.arn.org/ubb/ultimatebb.php?ubb=get_topic;f=13;t=001197. Accessed
on March 13, 2004.
[10] Dembski, William A. 1998. (a)
The Design Inference: Eliminating Chance through Small Probabilities.
Cambridge: Cambridge University Press.
(b) 1999. "Redesigning Science."
In W.A. Dembski, ed. Mere Creation: Science, Faith, and Intelligent Design.
Downers Grove, Ill.: InterVarsity Press.
(c) 1999. Intelligent Design:
The Bridge Between Science and Theology. Downers Grove, Ill.: InterVarsity
Press.;
(d) 2000. "The Third Mode of
Explanation: Detecting Evidence of Intelligent Design in the Sciences." In
W.A.Dembski, M. J. Behe, and S. C. Meyer, eds., Science and Evidence for
Design in the Universe. San Francisco: Ignatius Press.
(e) 2001. "What Intelligent
Design Is Not." In W. A. Dembski and J. M. Kushiner, eds., Signs of
Intelligence: Understanding Intelligent Design. Grand Rapids, Mich.: Brazos
Press. 2002.
(f) 2002. No Free Lunch: Why
Specified Complexity Cannot Be Purchased without Intelligence. Lanham, Md.:
Rowman and Littlefield.
[11] Dembski, "What Intelligent
Design Is Not."
[12] Perakh, Unintelligent
Design, chapter 1.
[13] Dembski, Intelligent
Design, and No Free Lunch.
[14] Dembski, Intelligent
Design, p. 159.
[15] Dembski, The Design
Inference, pp. 151154; No Free Lunch, pp. 115118.
[16] Dembski, The Design
Inference, pp. 7375.
[17] Plantinga, Alvin. 2001.
"Methodological Naturalism?" In R. T. Pennock, ed. Intelligent Design
Creationism And Its Critics. Cambridge: MIT Press, 349350.
[18] Dembski. "Redesigning
Science," p.107.
[19] (a) Perakh: (a1) "A
Consistent Inconsistency;" (a2) Unintelligent Design. (b).Young,
"Dembski's Explanatory Filter Delivers a False
Positive," and others.
[20] Ratzsch, Nature, Design, and Science. The
example of a false positive produced by the EF given in this book (pp. 166167)
is a case of driving on a desert road whose left side was flanked by a long
fence with a single small hole in it. A tumbleweed driven by wind happened to
cross the road in front of Ratzsch's car and rolled precisely through the sole
tiny hole. The event had an exceedingly small probability and was "specified" in
Dembski's sense (exactly as a hit of a bull'seye by an arrow in Dembski's
favorite example). Dembski's EF leads to the conclusion that the event in
question (tumbleweed rolling through the hole in the fence) was designed while
it obviously was due to chance; this is a false positive.
[21] Dembski. The Design
Revolution.
[22] Dembski, "Redesigning
Science," p. 111.
[23] Dembski. The Design
Inference, p. 136.
[24] Dembski. The Design
Inference, p. 17.
[25] Dembski. The Design
Inference, pp. 114115.
[26] (a) Nakaya, U. 1954. Snow Crystals: Natural and Artificial,
Cambridge, MA: Harvard University
Press. (b) Tape, W. 1994. Atmospheric Halos. Antarctic Research
Series, v. 69. American Geophysical Union.
[27] Dembski, No Free Lunch,
p. 12.
[28] Ibid., p. 127
[29] Gray, Robert M. 1991.
Entropy and Information Theory. Berlin: Springer Verlag.
[30] Dembski. No Free Lunch,
p. 166.
[31] Dembski. The Design
Inference, p. 94.
[32] Dembski. No Free Lunch,
p. 6.
[33] Ibid., pp. 159160.
[34] Ibid., pp. 166173.
[35] Ibid., p. 131.
[36] Landau, Lev D., and Evgeniy
M. Lifshits. 1971. Statisticheskaya physika (Statistical physics).
Moscow: Nauka; 40.
[37] Perakh. Unintelligent Design,
pp. 3435.
[38] Wolpert, David H and William G. Macready. 1997. "The No Free Lunch Theorems for
Optimization." IEEE Trans. Evol. Comp. v.1, no 1, 67–82.
[39] Dembski, No Free Lunch, pp. 200202.
[40] Dawkins, Richard. 1996. The Blind Watchmaker: Why the Evidence of Evolution
Reveals a Universe Without Design. New York: Norton.
[41] Chellapilla, Kumar and David B. Fogel. 1999. "CoEvolving Checkers
Playing Programs Using Only Win, Lose or Draw." SPIE's AeroSence '99:
Applications and Science of Computational Intelligence II (Orlando, Fla.:
5–9 April 1999).
[42] Altshuler, Edward E., and
Derek S. Linden. 1999. "Design of Wire Antennas Using Genetic Algorithms." In Y.
RahmatSamii and E. Michielssen, eds., Electromagnetic Optimization by
Genetic Algorithms (New York: Wiley): 211–248.
[43] Dembski. No Free Lunch,
pp. 224228.
[44] Dembski. 2002d. "Evolution's Logic of Credulity: An Unfettered
Response to Allen Orr."
www.designinference.com/documents/2002.12.Unfettered_Resp_to_Orr.htm. Accessed on December 26, 2002.
[45] Dembski. No Free Lunch, p. 203.
[46] Ibid., p. 182
[47] Perakh, 2004. "Reinventing
the Wheel." In Talk Reason,
http://www.talkreason.org/articles/wheel.cfm. Accessed on December 12, 2004.
[48] After this article was
submitted for publication, William Dembski published online three articles which
he claimed to be installments for the forthcoming set of papers under the
general title of Mathematical Foundation of Intelligent Design. One of them is
just a resurrected paper by Dembski of 1991, titled "Uniform Probability." The
other two can be seen at http://www.iscid.org/papers/Dembski_VariationalInformation_072404.pdf and at http://www.designinference.com/documents/2005.03.Searching_Large_Spaces.pdf.
These articles have been subjected to a strong critique revealing that the
mathematical exercises in these papers are irrelevant to intelligent design
"theory" and therefore these papers in no way can serve as a foundation of
intelligent design and of Dembski's concepts within that "theory's" framework.
The critique of Dembski's latest mathematically loaded papers can be seen, for
example, at http://www.talkreason.org/articles/math.cfm and http://www.talkreason.org/articles/newmath.cfm.
Mark Perakh is a
native of Ukraine. He had for many years been a professor of physics at various
universities in the USSR, as well as in three other countries. He came to the
US in 1978 as a visiting scientist at the IBM research center, later joined the
faculty of the California State University, and retired in 1994 as an emeritus. He has to his credit nearly 300 scientific
papers, four books and a number of patents, and was awarded a number of prizes
for his research including one from the Royal Society of London. He has also
been active in the anticreationismdebate (his
recent book is Unintelligent Design  Prometheus, 2004).
