Testability

Another frequent objection claims that ID does not yield testable or “falsifiable” predictions. This would be a damning objection to ID, if it were true. The ability of a theory to generate testable predictions—especially predictions that could, potentially, turn out false—is widely considered a sine qua non for scientific legitimacy. As discussed previously, however, design hypotheses do yield testable predictions that could, and someday might, be falsified by observations. In fact, curiously, another common objection to ID is that some design hypotheses already have been falsified. (See discussion of the “inept design” and “immoral design” objections, below.)

Testability and falsifiability can’t be evaluated as straightforwardly as many suppose, and neither criterion enables us to distinguish genuine science from pseudoscience. Moreover, the evidence for a theory doesn’t always come primarily from its predictive success, especially in so-called “historical” sciences like cosmology, paleontology, and evolutionary biology. We’ll revisit these issues in the last chapter. For present purposes, let’s grant that testability and falsifiability are hallmarks of a good scientific hypothesis.

In many cases, I suspect, this mistaken objection arises from conflating design arguments with design hypotheses. Arguments for design have received most of the attention in debates about the ID movement, so it is understandable that many critics of ID confuse design arguments with design hypotheses (or are simply unaware of the latter). The conclusion of a philosophical argument for design, such as the fine-tuning argument I presented earlier, doesn’t necessarily give us any prediction to test; but many design hypotheses do yield testable predictions.

There is, however, a subtler objection about testability that is worth addressing. Philosopher Elliott Sober contends that design hypotheses yield testable predictions only when conjoined (either explicitly or tacitly) with unwarranted assumptions about the designer. Here’s the gist of his argument. In order to test a design hypothesis, he says, we need to know whether it makes some possible observation more likely or less likely than rival hypotheses do.In a simple case with just two competing hypotheses, we can employ a rule known as the likelihood principle. The likelihood principle says that if an observation is more likely (more expected, or less surprising) assuming hypothesis H1 is true than if we assume H2 is true, then the observation counts as evidence favoring H1 over H2. For example, suppose we think a design hypothesis makes it more likely that we’ll find irreducibly complex structures inside cells, compared to the likelihood predicted by a rival evolutionary hypothesis. In that case, the observation that cells contain irreducibly complex structures would count as evidence favoring the design hypothesis over the evolutionary alternative. This principle is best understood within the context of Bayesian confirmation theory, which I briefly introduced in my discussion of the fine-tuning argument. For a fuller introduction to Bayesian confirmation theory, please refer to the chapters on Confirmation Theory and Evidence and Rationality in my online ebook Skillful Reasoning: An Introduction to Formal Logic and Other Tools for Careful Thought. The likelihood principle is introduced here. The problem, according to Sober, is that we cannot evaluate whether a design hypothesis makes an observation more or less likely unless we know what a designer is likely or unlikely to do. This is easy in the case of human designers, since we have abundant experience with human creativity. However, the designer responsible for instances of apparent design in nature would have to be a god or other supernatural being radically different from us, and we cannot predict what a supernatural designer would be likely or unlikely to do.

Without independent evidence about the designer’s goals and abilities, Sober contends, we can’t assess whether a design hypothesis makes any specific observation more likely or less likely than rival hypotheses do:

The problem is that the design hypothesis confers a probability … only when it is supplemented with further assumptions about what the designer’s goals and abilities would be if he existed. … There are as many likelihoods as there are suppositions concerning the goals and abilities of the putative designer. Which of these, or which class of these, should we take seriously?Elliott Sober, “The Design Argument,” in Neil A. Manson (ed.), God and Design: The Teleological Argument and Modern Science. (New York: Routledge, 2003), 38.

Consider, for example, the hypothesis that an intelligent designer created the first living organism on Earth. This hypothesis, by itself, tells us little about what we can expect to observe. For all the hypothesis says, the designer may have cared not a whit about the created organism’s progeny, and the creature might have died quickly before it could produce any offspring. Without making additional assumptions about the designer’s goals and abilities, the bare hypothesis that the first organism was designed doesn’t tell us what we are likely to observe. In order to test design hypotheses, therefore, the intelligent design theorist “needs independent evidence as to what the designer’s plans and abilities would be if he existed.”Elliott Sober, “The Design Argument,” in Neil A. Manson (ed.), God and Design: The Teleological Argument and Modern Science. (New York: Routledge, 2003), 38.

Sober thinks we have no evidence, independent of design hypotheses themselves, about a supernatural designer’s plans and abilities. I disagree: as a Christian, I think we have multiple lines of evidence about our Creator’s plans and abilities, including in particular those revealed through prophecies and testimonies in the Bible. Both young-earth and old-earth biblical creationists overtly conjoin scriptural evidence with empirical evidence when formulating and testing their hypotheses. However, this response to Sober’s objection will not help the ID movement. Proponents of ID claim that evidence of design can be found in nature using purely scientific methods, without relying on any religious sources. For this reason, they cannot answer Sober’s challenge by citing scriptural evidence.

How should ID theorists respond, then? From the standpoint of ID, Sober’s argument is vulnerable to refutation in several ways. I’ll summarize what I regard as its three most serious flaws, though I think it suffers from other problems as well.Here is a fourth problem, related to but separable from the others. In the quotation above, Sober asked which “suppositions concerning the goals and abilities of the putative designer” we should take seriously, insinuating that ID theorists can give no satisfactory answer to the question. The obvious answer is that we don’t have to make any such choice between differing assumptions or suppositions: we can simultaneously entertain multiple possibilities. Sober acknowledges this. (On page 39, he writes: “My claim is not that design theorists must have independent evidence that singles out a specification of the exact goals and abilities of the putative intelligent designer. They may be uncertain as to which of the goal-plus-abilities pairs GA1, GA2, … GAn is correct.”) He also concedes that ID advocates don’t have to determine exact probabilities, or even a precise range of probabilities, to test a design hypothesis. They only have to judge that the design hypothesis makes a specific observation somewhat more likely (or less likely) than rival hypotheses do. (On page 38, in the context of an example comparing a design hypothesis with a chance hypothesis, Sober writes: “What is required is not the specification of a single probability value, or even a range of such. All that is needed is an argument that shows that this probability is indeed higher than the probability that Chance confers on the observation.”) Nevertheless, he claims that ID theorists cannot even make such simple judgments unless they have independent evidence about the designer’s goals and abilities. Why not? Sober is not entirely clear on this point, but he seems to think that any such judgment would (in the absence of evidence) have to rely on arbitrary, unjustified assumptions about the designer’s goals and abilities.

He apparently fails to recognize that even if we remain agnostic about the designer’s goals and abilities, we can still estimate roughly how likely or unlikely we think an observation would be if a design hypothesis were true. As I mentioned previously, Bayesian probabilities are subjective probabilities, which represent degrees of belief. When we are maximally uncertain about some claim—that is, when we have not the slightest idea whether it is true or false—this uncertainty is represented with a probability of ½, or 50%. Clearly, we don’t have to assume anything about a designer’s goals or abilities in order to be completely uncertain whether, for example, he would be willing and able to create biological life. To the contrary, an attitude of uncertainty seems consummately reasonable when considering such questions independently of any scriptural or religious evidence.

Thus, the probabilities conferred by a design hypothesis need not be rooted in compelling evidence about the designer’s goals and abilities. They might instead reflect our uncertainty or paucity of evidence about those subjects. Even so, this would not prevent us from testing a design hypothesis. We can still make the relevant comparisons to see which hypothesis makes a specific observation more probable than its rivals do.
For one thing, his objection rests on the dubious claim that ID theorists cannot justify any assumptions at all about the capabilities or likely behavior of a supernatural designer. Behe, Meyer, and many other ID theorists have argued, to the contrary, that we have abundant evidence from our own experience about the effects of mind, or intelligent causation; and we have good reason to suppose that other minds—including supernatural ones—would share some similar capabilities. The very concept of a mind seems to include certain capacities relevant to evaluating design hypotheses, such as the capacity to produce functional information (specified complexity).

Secondly, even if Sober is right that such extrapolations from our own experiences are unjustified, it doesn’t follow that we are unable to test any design hypotheses. Some of the hypotheses we considered earlier yielded testable predictions without invoking additional assumptions about the designer. The predictions derived from Denton’s hypothesis, for example, do not seem to depend on any further assumptions about the designer’s goals or abilities. If the properties of biomatter are indeed finely-tuned for certain forms, as Denton speculates, this fine-tuning should be discoverable and even measurable (in principle) using methods analogous to those employed in cosmology and astrophysics. (Of course, the hypothesis presupposes that the designer was able to fine-tune the properties of matter, but this presupposition about the designer’s ability is built in to the hypothesis being tested. It is not an unmotivated or ad hoc auxiliary assumption.)

For further discussion of the Bayesian view, see the chapters on Confirmation Theory and Evidence and Rationality in my online ebook Skillful Reasoning: An Introduction to Formal Logic and Other Tools for Careful Thought.

A third problem with Sober’s argument has to do with the concept of testability itself. To his credit, Sober helpfully frames the testability issue in terms of Bayesian confirmation theory, the leading philosophical account of how scientific hypotheses are confirmed or disconfirmed by evidence. According to the Bayesian view, we test a hypothesis by making observations that provide evidence for it or against it. Often, this involves deriving some definite prediction from the hypothesis and then checking whether the prediction was correct.This testing strategy is known as the hypothetico-deductive method, which will be further discussed in Chapter 12 [coming soon]. See also this page from my Skillful Reasoning ebook. Testing needn’t always work that way, though. Sometimes, an observation that was not specifically predicted by a theory nonetheless provides good evidence for it; and, conversely, sometimes an observation that is compatible with a theory provides compelling evidence against it. Here are two simple, abstract examples to show how these counterintuitive situations can arise:

Suppose hypothesis H1 gives us no reason to think we’ll observe phenomenon P, nor any reason to think we won’t observe P. It is tempting to conclude that observing P would not provide any evidence at all, either for or against H1, since H1 neither predicts nor forbids P. However, the predictions of rival hypotheses must also be taken into account:

First example: Assume there is only one plausible alternative hypothesis, H2, which forbids P—that is, it predicts that P probably won’t occur. In other words, if we assume H2 is true, then P is extremely unlikely. In that case, observing P would drastically reduce our confidence in H2, thereby significantly increasing our confidence in H1. Thus, in this example, phenomenon P provides strong evidence for H1 even though it was not predicted by H1.

Second example: Assume instead that rival hypothesis H2 predicts P. In other words, if we assume H2 is true, then P is very likely to be observed. In that case, observing phenomenon P should increase our confidence in H2 at least to some degree, thereby decreasing our confidence in H1. So, in this second example, P provides evidence against H1 even though it is compatible with H1.

As the above examples illustrate, whether an observation counts as evidence for or against a hypothesis may depend crucially on the predictions of rival hypotheses. Sober’s argument does not take into account the role of rival hypotheses in theory testing, and this lacuna leaves room for a kind of testability he fails to consider. Even design hypotheses that yield no definite predictions might be testable in the Bayesian sense: specific observations could provide strong evidence either against them or for them, provided all plausibleBy “plausible,” in this context, I mean hypotheses with a prior probability that is not negligibly low. alternatives make those same observations very likely or very unlikely, respectively. Here’s a real-life case like that:

Consider again the hypothesis that an intelligent designer created the first living organism on Earth. Just to give it a name, let’s call it “D1”:

Hypothesis D1: The first living organism on Earth was designed.

This is a terribly uninformative design hypothesis. It tells us little about the designer’s abilities: the designer might be an all-powerful God or an advanced but finite extraterrestrial civilization. Nor does the hypothesis include information about the designer’s intentions and goals for the first organism, let alone for the organism’s progeny. Nevertheless, we may be able to test hypothesis D1, in the Bayesian sense (i.e., to determine whether an observation provides evidence for it or against it), by comparing its admittedly modest implications with the predictions given by rival hypotheses.

For example, the observations made by Douglas Axe, mentioned earlier in this chapter and in chapter 10,See discussion of the combinatorial problem in Chapter 10, which we revisited when considering the argument from specified complexity in this chapter. clearly have some evidential bearing on D1. As Stephen Meyer has argued, Axe’s experimental data suggests it is astronomically improbable that mere chance processes could have produced the amount of specified complexity required for the first organism.I summarized Meyer’s argument here. This means that Axe’s results were extremely surprising or unexpected—they had an extremely low subjective probability—from the perspective of the chance hypothesis, which is an important rival to D1. Indeed, some proponents of the chance hypothesis refused to believe Axe’s results, concluding instead that he must have made a mistake. Other proposed alternatives to the design hypothesis fare no better, Meyer argues. So, even if the design hypothesis gives us a modestly low degree of confidence that things should turn out the way Axe found them to be, this mediocre probability is vastly higher than the probabilities given by the most plausible rival hypotheses. Therefore, according to Bayesian confirmation theory, Axe’s observations constitute evidence for D1. (This is similar to the way in which cosmological fine-tuning discoveries provide evidence for design, as I argued previously: they drastically reduce our confidence in the most plausible alternatives.)

Other possible observations might provide evidence against D1. For example, a discovery of extraterrestrial microbes living on space dust might dramatically increase our degree of confidence in the “panspermia” hypothesis (which speculates that the first life on Earth arrived as microbes that evolved somewhere else in the galaxy), thereby decreasing our confidence in D1.The panspermia hypothesis isn’t strictly incompatible with design: God (or another designer) might have designed life somewhere else first and brought it here by way of space dust. For that matter, God might have designed the first organism by ordaining natural processes to produce it somewhere else in our galaxy. Nevertheless, insofar as panspermia is inconsistent with some other (more plausible) design hypotheses, a discovery of extraterrestrial microbes living on space dust would constitute evidence against the unspecific hypothesis that the first living organism on Earth was designed (in an unspecified way, by an unspecified designer). Thus, even though it is a relatively uninformative hypothesis, D1 still is amenable to testing—it can be confirmed or disconfirmed by evidence in the Bayesian way—without relying on arbitrary, unjustifed assumptions about the designer’s goals or abilities.

These examples show that the testability of a hypothesis depends not only on its own implications but also on the predictions of rival hypotheses. Regardless of what probability a hypothesis confers on some observation, that observation may or may not provide strong evidence for (or against) the hypothesis, depending on what probabilities rival hypotheses assign to the same observation.

There is much more to say on the topic of testability. We’ll return to it in chapter 12, where the concepts of testability, falsifiability, and theory confirmation will be examined in greater depth. For now, I hope this brief discussion suffices to show why the claim that design hypotheses are untestable does not hold up to scrutiny.