So, here we go:
First, Winston graciously acknowledges the eleP(T|H)ant:
I wrote here previously (“Design Detection in the Dark“) in response to blogger Elizabeth Liddle’s post “A CSI Challenge” at The Skeptical Zone. Now she has written a reply to me, “The eleP(T|H)ant in the room.” The subject of discussion is CSI, or complex specified information, and the design inference as developed by William Dembski.
CSI is the method by which we test chance hypotheses. For any given object under study, there are a variety of possible naturalistic explanations. These are referred to as the relevant chance hypotheses. Each hypothesis is a possible explanation of what happened to produce the object. If a given object is highly improbable under a giving hypothesis, i.e. it is unlikely to occur, we say that the object is complicated. If the object fits an independent pattern, we say that it is specified. When the object exhibits both of these properties we say that it is an example of specified complexity. If an object exhibits specified complexity under a given hypothesis, we can reject that hypothesis as a possible explanation.
The design inference argues from the specified complexity in individual chance hypotheses to design. Design is defined as any process that does not correspond to a chance hypothesis. Therefore, if we can reject all chance hypotheses we can conclude by attributing the object under study to design. The set of all possible chance hypotheses can be divided into hypotheses that are relevant and those that are irrelevant. The irrelevant chance hypotheses are those involving processes that we have no reason to believe are in operation. They are rejected on this basis. We have reason to believe that the relevant chance hypotheses are operating, and these hypotheses may be rejected by our applying the criterion of specified complexity. Thus, the design inference gives us reason to reject all chance hypotheses and conclude that an object was designed.
Originally, Liddle presented a particular graphic image of unknown origin and asked whether it is possible to calculate the probability of its being the product of design. In reply, I pointed out that knowing the potential chance hypotheses is a necessary precondition of making a design inference. Her response is that we cannot calculate the probabilities needed to make the design inference, and even if we could, that would not be sufficient to infer design.
He then enumerates what he thinks are my errors:
Elizabeth Liddle’s Errors
At a couple of points, Liddle seems to misunderstand the design inference. As I mentioned, two criteria are necessary to reject a chance hypothesis: specification and complexity. However, in Liddle’s account there are actually three. Her additional requirement is that the object be “One of a very large number of patterns that could be made from the same elements (Shannon Complexity).”
This appears to be a confused rendition of the complexity requirement. “Shannon Complexity” usually refers to the Shannon Entropy, which is not used in the design inference. Instead, complexity is measured as the negative logarithm of probability, known as the Shannon Self-Information. But this description of Shannon Complexity would only be accurate under a chance hypotheses where all rearrangement of parts are equally likely. A common misconception of the design inference is that it always calculates probability according to that hypothesis. Liddle seems to be plagued by a vestigial remnant of that understanding.
Um, no – it’s Dembski himself and his followers that seem to be plagued by that vestigial remnant. Perhaps they should have it surgically removed. In a recent talk (24th January 2013) , which I partially transcribed, Dembski said:
Well, it turned out what was crucial for detecting design was this what I called Specified Complexity, that you have a pattern, where the pattern signifies an event of low probability, and yet the pattern itself is easily described, so it’s specified, but also low probability. And I ended up calling it Specified Complexity, I don’t want to get into the details because this can be several lectures in itself, but so there was this marker, this sign of intelligence, in terms of specified complexity, it was a well-defined statistical notion, but it turned out it was also connected with various concepts in information theory, and as I developed Specified Complexity, and I was asked back in 05 to say what is the state of play of Specified Complexity, and I found that when I tried to cash it out in Information Theoretic terms, it was actually a form of Shannon Information, I mean it had an extra twist in it, basically it had something called Kolmogorov complexity that had to be added to it.
Actually, Dembski is being a little curly here. In his 2005 paper, he says:
It’s this combination of pattern- simplicity (i.e., easy description of pattern) and event-complexity (i.e., difficulty of reproducing the corresponding event by chance) that makes the pattern exhibited by ( ψR) — but not (R) — a specification.
The “complexity” part of Specified Complexity is the “event-complexity” – P(T|H), which can only be the part referred to as “Shannon information” in his talk, Moreover, in his 2005 paper, his worked examples of “event-complexity” are of Shannon entropy, although he does indeed say that when computing P(T|H), H must be “the relevant chance hypothesis”, which of course will not necessarily be random independent draw. But not only does Dembski spend no time showing us how to compute P(T|H) for anything other than random independent draw (i.e. the Shannon entropy), he goes to great lengths to show us examples where P(T|H) is random independent draw, the negative log of which gives us Shannon entropy. And the only attempts to calculate any version of CSI (including various versions where the Specification is function) that I have seen have used random independent draw.
Moreover, the “specified” part of CSI is the “pattern-simplicity” part – the Kolmogorov part. This is not Kolmogorov complexity but its negation – the more compressible (less Kolmogorov complex) a pattern is, the more “specified” it is. Dembski must be aware of this – perhaps he thought his 2013 audience would be happy to accept that “Kolmogorov complexity” has something to do with the Complexity in in Specified Complexity, even though the Kolmogorov part is is the Specification part, where more Specified = less complexity, and that “Shannon information” must be something to do with the specification (a sequence that means something, right?), whereas in fact that’s the “complexity” (entropy) part. So he’s doing some voodoo here. In the 2005 paper, Dembski quite clearly defines Specified Complexity as a sequence that improbable under some hypothesis (possibly random independent draw, in which case the answer would be “cashed out” in Shannon information) i.e. “complex”, AND is readily compressible (has low Kolmogorov complexity, and therefore specified). I was recently taken to task by Barry Arrington for violating language by using Dembski’s terminology. heh. I think Dembski himself is exploiting his own language violations to squirt ink, here.
But I do accept that the eleP(T|H)ant is there, in the 2005 paper, in the small print. Indeed, I pointed it out. It is not my vestigial organ that is causing the problem. It’s precisely that vestigial organ I am calling the eleP(T|H)ant, and asking to be addressed.
As I emphasized earlier, the design inference depends on the serial rejection of all relevant chance hypotheses. Liddle has missed that point. I wrote about multiple chance hypotheses but Liddle talks about a single null hypothesis. She quotes the phrase “relevant [null] chance hypothesis”; however, I consistently wrote “relevant chance hypotheses.”
Fine. I am all for serial rejection of specific null hypotheses. It’s how science is usually done (although it has its problems). In which case we all continue to ignore the single eleP(T|H)ant considered rejectable by Dembski in one single equation and get on with doing some science.
Liddle’s primary objection is that we cannot calculate the P(T|H), that is, the “Probability that we would observe the Target (i.e. a member of the specified subset of patterns) given the null Hypothesis.” However, there is no single null hypothesis. There is a collection of chance hypotheses. Liddle appears to believe that the design inference requires the calculation of a single null hypothesis somehow combining all possible chance hypotheses into one master hypothesis. She objects that Dembski has never provided a discussion of how to calculate this hypothesis.
Yes, indeed I do object. Like you, Winston, I entirely agree that “there is no single null hypothesis”. That’s why Dembski’s 2005 is so absurd – it implies that we can compute a pantechnicon null to represent “all relevant chance hypotheses” and plug the expected probability distribution under that null into his CSI formula. Of course it can’t be done. You know it, and I expect that Dembski, being no fool, knows it too (just as I am sure he knows that “pattern simplicity” was a tactical error, hence his attempt to imply that “Kolmogorov complexity” is part of the CSI recipe, even though the Kolmogorov part is the simplicity part, and the “complexity” part is simple improbability, with equal value for a meaningful pattern and for the same pattern drawn “by chance”).
But that is because Dembski’s method does not require it. Therefore, her objection is simply irrelevant.
um, Dembski’s method does require it. Nothing in that 2005 paper implies anything other than a single null. Sure, that null is supposed to, be H, namely
…the relevant chance hypothesis that takes into account Darwinian and other material mechanisms
But I see only one H. Are you sure you are not seeing pink eleP(T|H)ants, Winston?
Can we calculate the probabilities required to reject the various chance hypotheses? Attempting to do so would seem pretty much impossible. What is the probability of the bacterial flagellum under Darwinian evolution? What is the probability of a flying animal? What is the probability of humans? Examples given of CSI typically use simple probability distributions, but calculating the actual probabilities under something like Darwinian evolution is extremely difficult.
Yes indeed. Thank you, Winston.
Nevertheless, intelligent design researchers have long been engaged in trying to quantify those probabilities. In Darwin’s Black Box, Mike Behe argues for the improbability of irreducibly complex systems such as the bacterial flagellum. In No Free Lunch, William Dembski also offered a calculation of the probability of the bacterial flagellum. In “The Case Against a Darwinian Origin of Protein Folds,” Douglas Axe argues that under Darwinian evolution the probability of finding protein folds is too low.
Yes, IC is probably the best argument for ID. Unfortunately, it’s still terrible. Dembski and Behe both made their bacterial flagellum calculations (well, assertions – I don’t see much in the way of calculations in either book) without taking into account at least one possible route, i.e. that proposed by Pallen and Matzke. Pallen and Matzke’s route may well be wrong, but when computing probabilities, at the very least, the probability of such a route should be taken into account. Much more to the point, as Lenski and his colleagues have shown (both in their AVIDA program and in their E.coli cultures) being “IC” by any of the definitions given by Behe or Dembski is not a bar to evolvability, either in principle (AVIDA) or practice (e-coli). Dembski writes, in No Free Lunch:
The Darwinian mechanism is powerless to produce irreducibly complex systems for which the irreducible core consists of numerous diverse parts that are minimally complex relative to the minimal level of function they need to maintain.
This claim has simply been falsified. Of course it is still possible that the bacterial flagellum somehow couldn’t have, or didn’t, evolve. But the argument that “the Darwinian mechanism” is in principle “powerless” to evolve an IC system is now known to be incorrect. As for Axe’s paper, essentially his claim is the same, namely that functional protein folds are deeply IC (using Behe’s IC-pathway definition), and are therefore improbable by Darwinian mechanisms. But he infers this via a similarly same falsified claim:
So, while it is true that neither structure nor function is completely lost along the mutational path connecting the two natural sequences, natural selection imposes a more stringent condition. It does not allow a population to take any functional path, but rather only those paths that carry no fitness penalty.
It turns out there is no such constraint. “Natural selection” as we now know, from both empirical observations and mathematical models is merely a bias in the sampling, in any one generation, of favour of traits that increase the probability of reproductive success in the current environment. As most mutations are near-neutral, many will become highly prevalent in the population, including some substantially deleterious ones (ones that “carry a fitness penalty”. So again, the argument that IC structures are unevolvable, or only evolvable by pathways that do not include deleterious steps, is simply unfounded. Falsified, in the strict Popperian sense.
Again, moving on:
While it is unreasonable to calculate the exact probabilities under a complex chance hypothesis, this does not mean that we are unable to get a general sense of those probabilities. We can characterize the probabilities of the complex systems we find in biology, and as the above research argues those probabilities are very small.
It is indeed “unreasonable to calculate the exact probabilities under a complex chance hypothesis”. Precisely. But I’d go further. It is even more unreasonable to make a rough guesstimate (I hate that word, but it works here). The entire concept of CSI is GIGO. If you start with the guesstimate that Darwinian mechanisms are unlikely to have resulted in T, then P(T|H) will be, by definition, small. If you start with the guesstimate that Darwinian mechanisms could well have resulted in T, then P(T|H) will be, by definition, large. Whether the output from your CSI calculation reaches your threshold (ah, yes, that threshold….) for rejection therefore depends entirely on the Number You First Thought Of.
And so, indeed, to get a defensible Number, we must turn to actual empirical research. But the only “above research” cited is Behe’s (scarcely empirical, rebutted by Pallen and Matzke, and contains no actual estimate), and Axe’s, which needs to be taken together with the vast number of other empirical research into the evolution of protein folds. And both are based on a falsified premise. Sure one might be able to get “get a general sense of those probabilities”, but in order to make the strong claim that non-design must be rejected (at p<10^-150!) we need something more than a “general sense of those possibilities”. Bear in mind that no scientist rejects ID, at least qua scientist. There is no way to falsify it (how would you compute the null?) ID could well be correct. But having a hunch that non-design routes can’t do the job is not enough for an inference with that kind of p value. No sir not nohow.
In summary so far: Ewert has done little more than agree with me that we cannot reject non-design on the basis of the single pantechnicon null as proposed in Dembski’s 2005 paper, and would therefore presumably also agree that quantity chi is uncomputable, and therefore useless as a test for design (which was the point of my original glacier puzzle). Instead he argues that research does suggest that Darwinian mechanisms are implausible, but does so citing research that is based on the assumption that Darwinian mechanisms cannot either produce IC systems (they can) or by IC pathways (they can) or by IC pathways that include deleterious steps (they can). Now it remains perfectly possible that bacterial flagella and friends could not have evolved. But the argument that they cannot have evolved because “natural selection… does not allow a population to take any functional path, but rather only those paths that carry no fitness penalty” fails, because the premise is false.
Earman and Local Inductive Elimination
Above, I mentioned the division of possible chance hypotheses into the relevant and irrelevant categories. However, what if the true explanation is an unknown chance hypothesis? That is, perhaps there is a non-design explanation, but as we are ignorant of it, we rejected it along with all the other irrelevant chance hypotheses. In that case, we will infer design when design is not actually present.
Dembksi defends his approach by appealing to the work of philosopher of physics John Earman, who defended inductive elimination. An inductive argument gives evidence for its conclusion, but stops short of actually proving it. An eliminative argument is one that demonstrates its conclusion by proving the alternatives false rather than proving the conclusion true. The design inference is an instance of inductive elimination: it gives us reason to believe that design is the best explanation.
Liddle objects that Dembski is not actually following what Earman wrote, and she quotes from Earman: “Even if we can never get down to a single hypothesis, progress occurs if we succeed in eliminating finite or infinite chunks of the possibility space. This presupposes of course that we have some kind of measure, or at least topology, on the space of possibilities.”
Dembski has not defined any sort of topology on the space of possibilities. He has not somehow divided up the space of all possible hypotheses and systematically eliminated some or all. Without that topology, Dembski cannot claim to have eliminated all the chance hypotheses.
True. But nor can he claim to have eliminated any unless he actually calculates the ele(P(T|H)ant for each.
However, Liddle does not appear to have understood what Earman meant. Earman was not referring to a topology over every conceivable hypothesis, but over the set of what we might call plausible hypotheses. In Earman’s approach, inductive elimination starts by defining the set of plausible hypotheses. We do not consider every conceivable hypothesis, but only those hypotheses which we consider plausible. Only then do we define a topology on the plausible hypotheses and work towards eliminating the incorrect possibilities.
Hang on. I venture to suggest that Winston has not understood what Earman meant. Let me write it out again, with feeling:
Even if we can never get down to a single hypothesis, progress occurs if we succeed in eliminating finite or infinite chunks of the possibility space. This presupposes of course that we have some kind of measure, or at least topology, on the space of possibilities.
In other words, surely: you can only partition the total possibility space into plausible and implausible sections IF you first have “some kind of measure, or at least topology, on the space of possibilities.” Then of course, you can pick off the plausibles one by one until you are done. So my objection stands. And if that isn’t what Earman meant, then I have no idea what he meant. It seems pretty clear from his gravity example that that’s exactly what he meant – entire sections of hypothesis space could be eliminated first, leaving some plausibles.
In his discussion of gravitational theories, Earman points out that the process of elimination began by an assumption of the boundaries for what a possible theory would look like. He says:
Despite the wide cast of its net, the resulting enterprise was nevertheless a case of what may properly be termed local induction. First, there was no pretense of considering all logically possible theories.
Later, Earman discusses the possible objection that because not all logically possible theories were considered, it remains possible that true gravitational theory is not the one that was accepted. He says:
I would contend that all cases of scientific inquiry, whether into the observable or the unobservable, are cases of local induction. Thus the present form of skepticism of the antirealist is indistinguishable from a blanket form of skepticism about scientific knowledge.
In contrast to Liddle’s understanding, Earman’s system does not require a topology over all possible hypotheses. Rather, the topology operates on the smaller set of plausible hypotheses. This is why the inductive elimination is a local induction and not a deductive argument.
I think Winston is confused. Certainly, Earman does not require that all hypotheses be evaluated, merely those that are plausible. Which is fine. However he DOES require that first the entire set of possible hypotheses be partitioned into plausible and implausible, so that you can THEN “operate..on the smaller set of plausible hypotheses”.
So we could, for instance, eliminate at a stroke, all theories that require, say, violation of the 2nd Law of thermodynamics; or hitherto undiscovered fundamental forces; and material designers, as that would only push back the problem as to where they came from, and already we are short of cosmic time. We don’t need to even worry about that herd of ele(P(T|H)ants. And as Darwinian processes require self-replicators, we can eliminate Darwinian processes as an account of the first self-replicators. And so on. But that still leaves “Darwinian and other material processes” as Dembski correctly states in his 2005 paper. And those null distributions have to be properly calculated if they are going to be rejected by null hypothesis testing. Which is not of course what Earman is even talking about – his entire book is about Bayesian inference.
Which brings me to my next point: Dembski’s 2002 piece, cited by Ewert below, precedes his 2005 paper, and indeed, No Free Lunch, and presents a Bayesian approach to ID, not a Fisherian one. In the 2005 paper, Dembski goes to excruciating length to justify a Fisherian approach, and explicitly rejects a Bayesian one:
I’ve argued at length elsewhere that Bayesian methods are inadequate for drawing design inferences. Among the reasons I’ve given is the need to assess prior probabilities in employing these methods, the concomitant problem of rationally grounding these priors, and the lack of empirical grounding in estimating probabilities conditional on design hypotheses.
Indeed they do. But Fisherian methods don’t get him out of that responsibility (for grounding his prior), they merely allow him to sneak them in via a Trojan EleP(T|H)ant.
Furthermore, Dembski discusses the issue in “Naturalism’s Argument from Invincible Ignorance: A Response to Howard Van Till,” where he considers the same quote that Liddle presented:
In assessing whether the bacterial flagellum exemplifies specified complexity, the design theorist is tacitly following Earman’s guidelines for making an eliminative induction work. Thus, the design theorist orders the space of hypotheses that naturalistically account for the bacterial flagellum into those that look to direct Darwinian pathways and those that look to indirect Darwinian pathways (cf. Earman’s requirement for an ordering or topology of the space of possible hypotheses). The design theorist also limits the induction to a local induction, focusing on relevant hypotheses rather than all logically possible hypotheses. The reference class of relevant hypotheses are those that flow out of Darwin’s theory. Of these, direct Darwinian pathways can be precluded on account of the flagellum’s irreducible and minimal complexity, which entails the minuscule probabilities required for specified complexity. As for indirect Darwinian pathways, the causal adequacy of intelligence to produce such complex systems (which is simply a fact of engineering) as well as the total absence of causally specific proposals for how they might work in practice eliminates them.
Dembski is following Earman’s proposal here. He defines the boundaries of the theories under consideration. He divides them into an exhaustive partition, and then argues that each partition can be rejected therefore inferring the remaining hypothesis, design. This is a local induction, and as such depends on the assumption that any non-Darwinian chance hypothesis will be incorrect.
Yep, he is. And he is using Bayesian reasoning: his rejection of indirect Darwinian pathways is rejected because of his higher priors on something else, namely “the causal adequacy of intelligence to produce such complex systems” (which I would dispute, but let’s say per arguendo he is justified); and on “the total absence of causally specific proposals“, which is ludicrous. We do not demand to know the causally specific pathway by which a rock descended from a cliff before we can infer that erosion plus gravity were the likely cause. What we need to know is whether indirect Darwinian pathways can produce “IC” systems, and we know they can. It is “causally adequate” (and a heck sight more “causally adequate” than an “intelligence” with no apparent physical attributes capable of assembling a molecule – intelligence doesn’t make things, intelligent material beings do).
But all that is moot, because his 2002 is a Bayesian inference, not a Fisherian one. If he wants to play the Fisherian game, then he’s welcome to do it, but in that case he has to carefully construct the probability distribution of every null he wants to reject, and, if he succeeds, claim only to have rejected that null, not a load of other nulls he thinks he can get away with. Which then won’t allow him to reject “non-design”. Merely the nulls he’s modelled.
Frankly, if I were an IDer I’d go down the Bayesian route. But, as Dembski says, it does present problems.
At the end of the day, the design inference is an inductive argument. It does not logically entail design, but supports it as the best possible explanation.
“Best possible explanation” is a Bayesian inference, not a Fisherian one. Fine. But let’s bury CSI in that case.
It does not rule out the possibility of some unknown chance hypotheses outside the set of those eliminated.
It does not even rule out hypotheses within the set of those claimed to have been eliminated, because you can’t do that unless you can compute that null. Nothing Winston has written here gets him, or Dembski, off the hook of having to reject a Darwinian null, for a post-OoL system, because Darwinian process pass any plausibility test. So if you want to reject it, it you have to compute it. Which is impossible. It would be like trying to reject the null of that today’s weather was a result of set of causal chains X, where X is a set of specific hypothetical turbulent states, and then concluding that today’s weather wasn’t a result of a turbulent state. All we can say is that we know that turbulence leads to striking but unpredictable weather systems, and so we can’t reject it as a cause of this one.
However, rejecting the conclusion of design for this reason requires the willingness accept an unknown chance hypothesis for which you have no evidence solely due to an unwillingness to accept design. It is very hard to argue that such a hypothesis is actually the best explanation.
Indeed. But there is sleight of hand here. Darwinian processes is a known “chance hypothesis” (Dembski’s and your term, not mine), but the specific causal chain for any one observed system is probably unknown. That doesn’t mean we can ignore it, any more than we can reject ID because there are many possible ID hypotheses (front-loading; continuous intervention; fine-tuning; OoL only, whatever). And indeed there is a double standard. If we deem “intelligence” “causally adequate to produce such complex systems” because it is “is simply a fact of engineering”, despite the fact that we have no hypothesis for any specific causal pathway by which a putative immaterial engineer could make an IC system, then we can at the very least deem Darwinian processes “causally adequate” because we know that those processes can produce IC systems, by IC pathways that include deleterious steps.
Liddle objects that we cannot calculate the probability necessary to make a design inference. However, she is mistaken because the design inference requires that we calculate probabilities, not a probability.
Well, there was slightly more than to my objection than the number of hypotheses! And if Dembski agrees there must be more than one, I suggest that he retracts CSI as defined in his 2005 paper.
Each chance hypothesis will have it own probability, and will be rejected if that probability is too low. Intelligent design researchers have investigated these probabilities.
I have not seen a single calculation of such a probability that did not rest on the falsified premise that IC systems cannot evolve by indirect pathways, including those that include deleterious steps.
Liddle’s objections to Dembski’s appeal to Earman demonstrate that she is the one not following Earman. Earman’s approach involves starting assumptions about what a valid theory would look like, in the same way that any design inference makes starting assumptions about what a possible chance hypotheses would look like.
It also involves Bayesian inference, which Dembski specifically rejects. It also seems to involve rejecting a perfectly plausible hypothesis on falsified grounds.
In short, neither of Liddle’s objections hold water. Rather both appear to be derived from a mistaken understanding of Dembski and Earman.
Well, no. They remain triumphantly watertight; Winston has largely conceded their validity, by de facto agreeing with me that Dembski’s definition of “chi” is useless, and thus his Fisherian rejection method invalid. Winston substitutes a Bayesian approach, in which he improperly eliminates perfectly plausible hypotheses (e.g. indirect Darwinian pathways for IC systems), and offers no method for calculating them, merely appealing to researchers who have failed to note that IC is not an impediment to evolution. He appears himself to have misunderstood Earman, and far from my “misunderstanding” Dembski, Winston appears to be telling us that Dembski did not say what he categorically did say in 2005. At any rate, if Dembski in his 2005 was really telling us that we needed to use a Bayesian approach to the Design Inference and that multiple nulls must be rejected in order to infer Design, then it’s odd that his words appear to indicate the precise opposite.