Commentary on William A. Dembski's "No Free Lunch: Why Specified Complexity Cannot Be Purchased without Intelligence"

(Lanham, Maryland:Rowman & Littlefield Publishers, Inc., 2002, 404pp.)

Review commentary by Wesley R. Elsberry

Capsule review - a first look at "No Free Lunch"

[Note: Some folks apparently object to me having made comments so quickly after receiving my copy of NFL. I'll note that the positive comments seemed to appear just as quickly, but they haven't been dismissed on those grounds. I am interested in presenting accurate information. I will correct any inaccuracies and credit the person making the correction. Just send email to my "welsberr" account on inia.cls.org]

"No Free Lunch" and non-payment on a promissory note

"No Free Lunch" brings us up to date with Dembski's thoughts on evolution and intelligent design. For those who are proponents of intelligent design antievolution, this book will be a must-read, as Dembski is the foremost philosopher for the ID movement. The first two chapters give a summary of Dembski's framework for making "design inferences". Those who have read Dembski's monograph, "The Design Inference", will see much that is familiar -- and a few things that have changed a bit. This is where we learn about "specified complexity" and what makes a "specification" a good thing in inferring design. The third chapter makes an argument that specified complexity and information theory are closely related, ending up with Dembski's assertion of a fourth law of thermodynamics, a conservation law for "complex specified information" (CSI). Chapter four moves on to Dembski's take on evolutionary algorithms, those computational analogues of biological natural selection. This is where Dembski deploys Wolpert and MacReady's "No Free Lunch" results to argue that evolutionary algorithms are incapable of producing the CSI we might identify in biological systems. Chapter five takes us to a discussion of Michael Behe's concept of "irreducible complexity" (IC), asserting relationships between specified complexity and IC, and offering a new, improved version of IC to boot. Chapter six gives us Dembski's view of how his notion of specified complexity in particular and intelligent design in general can be usefully deployed as a scientific research program.

Throughout the book, Dembski takes up arguments that critics have made. As one of those critics, I can offer my opinion that Dembski fails to track those critical arguments in certain crucial respects.

For instance, Dembski brushes off a criticism concerning the reliability of his "explanatory filter" by noting that the objection is the problem of induction, but fails to either solve the problem of induction or retract the claim of reliability. That's philosophical humor, by the way. Dembski is not going to solve the problem of induction. That means that he should have retracted his claim of reliability. Just to be clear, let's see what Dembski means by saying that his Explanatory Filter/Design Inference/Specified Complexity criterion is reliable.

I want, then, to argue that specified complexity is a reliable criterion for detecting design. Alternatively, I want to argue that the complexity-specification criterion successfully avoids false positives -- in other words, whenever it attributes design, it does so correctly.

-- WA Dembski, "No Free Lunch", p.24

The above is not a typical statement for "scientific inquiry". It describes the operation of an oracle, not an inference.

Some may object that "success" need not refer to the 100% reliability that Dembski's words above seem plainly to invoke. But we have further testimony from Dembski that that is exactly what is meant.

[...] Biologists worry about attributing something to design (here identified with creation) only to have it overturned later; this widespread and legitimate concern has prevented them from using intelligent design as a valid scientific explanation.

Though perhaps justified in the past, this worry is no longer tenable. There now exists a rigorous criterion complexity-specification for distinguishing intelligently caused objects from unintelligently caused ones. Many special sciences already use this criterion, though in a pre-theoretic form (e.g., forensic science, artificial intelligence, cryptography, archeology, and the Search for Extra-Terrestrial Intelligence). The great breakthrough in philosophy of science and probability theory of recent years has been to isolate and make precise this criterion. Michael Behe's criterion of irreducible complexity for establishing the design of biochemical systems is a special case of the complexity-specification criterion for detecting design (cf. Behe's book Darwin's Black Box).

What does this criterion look like? Although a detailed explanation and justification is fairly technical (for a full account see my book The Design Inference, published by Cambridge University Press), the basic idea is straightforward and easily illustrated. [...]

-- W.A. Dembski, "Science and design", First Things, Oct. 1998, http://www.firstthings.com/ftissues/ft9810/dembski.html, last accessed 2002/01/20.

Further, Dembski has never bothered to propose an effective empirical test methodology for his Explanatory Filter. Instead, it has been left to critics like myself to propose empirical methods of determining whether Dembski's claims of reliability have any grounding in fact.

Dembski has, so far, not analyzed potential counterexamples. I proposed at Haverford College last June that Dembski "do the calculation" for the Krebs citric acid cycle and the impedance-matching apparatus of the mammalian middle ear. Dembski has not done so.

In other places, Dembski fails to take up the arguments of critics, as in Dembski's mischaracterization of a program written by Richard Dawkins. Two out of three of the steps that Dembski says characterize the program are, in fact, Dembski's own invention, appearing nowhere in Dawkins's work. The sad thing is that criticism of precisely this point was made by me in email to Dembski back in October of 2000. It would have been easy for Dembski to fix, but it did not happen.

The most disappointing aspect of "No Free Lunch", though, has to do with section 5.10, "Doing The Calculation". Dembski had promised, under critical questioning, to publish an example of the application of his framework for inferring design from "The Design Inference" as it would be applied to a non-trivial example of a biological system. Section 5.10 is apparently what Dembski intended to serve as payment on that promissory note. However, it fails to deliver on several points. Dembski does not establish that the example, that of a bacterial flagellum, has a specification according to the usage in "The Design Inference". Dembski also fails to enumerate and then eliminate multiple relevant chance hypotheses, as indicated in "The Design Inference". Dembski especially does not evaluate the hypothesis that the bacterial flagellum developed through evolutionary change; a curious omission given the context. The single "chance" hypothesis that Dembski does bother to consider is a marginal refinement on the old antievolution standby, "random assembly". At least, the technical jargon looks denser around Dembski's argument than I've seen around "tornado in a junkyard" presentations. But all in all, section 5.10 does little to help those who wanted to see how a design inference could be rigorously applied to biological examples.

This book brings together in one place many of Dembski's ideas. The arguments range from very technical mathematical and logical formulations to pop culture examples to illustrate points. In my opinion, the book suffers from some very deep flaws -- it seems to me that Dembski misuses the NFL results to argue against the capacity of evolutionary algorithms to solve problems. Many more minor problems are apparent to me in even my brief acquaintance with the book. Still, if one wished to buy one book of Dembski's rather than the four that now grace my bookshelf, "No Free Lunch" would be the one to get.


Second draft, 1000 word limit

I've been a long-time critic of William A. Dembski's ideas, ever since we presented talks at the 1997 "Naturalism, Theism, and the Scientific Enterprise" conference. I've previously published a review of his book, "The Design Inference" in Reports of the National Center for Science Education, and I've coauthored a peer-reviewed paper which appeared in the November 2001 issue of the journal "Biology And Philosophy" on the topic of Dembski's "explanatory filter". I've been awaiting the publication of Dembski's "No Free Lunch" ever since he announced that he was working on it back in 1999. Dembski had indicated that this book would answer a number of my criticisms of his ideas.

My first impressions of the book, though, have not been positive. The interesting thing about this book is how it brings together in one volume so many false assertions and flawed arguments. The problems range from the trivial to the profound. However, if one is looking for a compact guide to Dembski's thought, this is the volume to get.

The most profound problem in Dembski's "No Free Lunch" is why evolutionary algorithms and biological evolution should be required to "generate" specified complexity, with emphasis on Dembski's narrow connotation of "generate". Fitness functions and environments represent rich sources of information; algorithms and evolutionary processes which alter bit strings and genomes in an iterated interaction may not "generate" specified complexity in Dembski's sense of the term, but it remains unclear that this excludes biological evolution from accounting for the diversity and complexity of organisms derived from a common ancestor.

My vote for the next most profound problem in Dembski's "No Free Lunch" is that while the subtitle says, "Why specified complexity cannot be purchased without intelligence", Dembski seems not to offer any coherent account of how specified complexity *can* be purchased *with* intelligence. Dembski seems to treat this as a "brute given". Dembski argues that algorithms and natural law cannot "generate" specified complexity. By implication, any logical calculus that could underwrite rationality in humans or "unembodied designers" is also incapable of "generation" of specified complexity in an intelligent agent. This leaves us with only irrational processes in intelligent agents as possible means of "purchasing" specified complexity, if we accept Dembski's arguments and assertions.

A major fault with Dembski's "No Free Lunch" concerns what did not appear within its pages. Dembski has been promising critics, myself included, that he would present an example of application of his "rigorous" complexity-specification criterion, as given in "The Design Inference", to one of the "complex, information-rich structures of biology". Critics have argued that Dembski's framework is too unwieldy to apply to non-trivial real world problems and systems; what better way to prove them wrong than if Dembski publicly applied the full "design inference" regimen to a biological example, especially since he was on record as saying that this work had already been done. Dembski made that claim in a 1998 article, so why we had to wait over two years for an attempt to be published makes an intriguing question. In section 5.10 of "No Free Lunch", Dembski does discuss a bacterial flagellum in terms of probabilities, but it fails to live up to the prior billing. While describing the mechanism of formulating and justifying a "specification" took up a whole chapter of Dembski's "The Design Inference", he used a single paragraph in "No Free Lunch" to make a simple declaration that a "specification is never a problem" for biological examples. This begs the question of whether we are dealing with a "good pattern" (or specification), or a "bad pattern" (or fabrication) for inferring "design". The remainder of the example deals with establishing the probability of building a bacterial flagellum by chance. Dembski works up precisely one "chance" hypothesis, though his "design inference" clearly states that one must generate and eliminate all the "relevant" chance hypoteheses in order to conclude "design". The "chance" hypothesis we see looks somewhat more technical than the common "random assembly" antievolution argument, but not qualitatively so. Notably, Dembski fails to develop and evaluate a hypothesis based upon biological descent with modification; an odd omission given the context. The choice of example was poor, since I pointed out to Dembski that an effective test of his framework could only occur where we had good evidence that the system being analyzed was, in fact, due to an evolutionary process. I even suggested to him that analysis of the Krebs citric acid cycle or the impedance-matching apparatus of the mammalian middle ear would be good candidates. Instead, Dembski analyzed the flagellum of E. coli, where we have no fossil record or definitive molecular evidence concerning its origin. All in all, this was a promissory note that was not made good.

Among the more minor errors is Dembski's 3-step description of Richard Dawkins's "weasel" program from "The Blind Watchmaker". 2 out of the 3 steps Dembski gives are Dembski's own inventions, appearing nowhere in Dawkins's work. Sadly, Dembski chose to ignore my email and publicly posted critique from October, 2000 pointing out this problem. Ironically, Dembski offers a "more realistic" modification of Dawkins's program, which coincidentally comes much closer to what Dawkins described. Elsewhere, Dembski insists on an over-broad connotation of "evolutionary algorithms". Previously, Dembski had stated that "neural nets" were instances of evolutionary algorithms; now, Dembski says that "training neural nets" are instances of evolutionary algorithms. The previous claim was simply false and the new claim is based upon the fact that some people do apply evolutionary computation to the problem of training neural nets. It reduces to the claim that instances of evolutionary computation are evolutionary computation; mentioning "neural nets" at all in that context seems unlikely to do anything but lead readers to the erroneous conclusion that the original claim has not been abandoned. Such basic errors as these reduce the effectiveness of his argumentation.

Despite the promotional hype, "No Free Lunch" is not a work that will disestablish evolutionary biology. Its arguments, especially where they touch upon evolutionary algorithms and actual biology, betray a superficiality of acquaintance with those fields. It is an important book, however, in the sense that it represents the best that the "Intelligent Design" movement has to offer. I am a critic of Dembski's ideas because he can be counted upon to make the most interesting errors of any of the current "Intelligent Design" proponents. But for those who are looking for a "magic bullet" to oppose evolution, the ammo of "No Free Lunch" is a dud.


Early version submitted to Barnes & Noble