Tuesday, July 10, 2012

Subjective specification makes CSI non-functional

It has been claimed a number of times that the specification in complex specified information (CSI) is subjective, at least to some degree.  The claim is that even though it is possible to quantify specification, to some degree, it is still at some level dependent upon a human domain of discourse for its description.

However, if it is true that specification is always subjective this presents a problem.  To understand why, we must examine the problem CSI is attempting to solve.

In bygone days, creationists used to argue that the universe must have been designed because many parameters are so finely tuned, and the combination of so many independent finely tuned parameters is vastly improbable.  While intuitively this seems like a knock down drag out argument, there is a logical fallacy at work.

Take a jar of multicolor marbles.  It happens to be configured in such a way that the marbles spell out the word "Supercalifragilisticexpialidocious" around the edge of the jar.  However, mathematically speaking, this particular configuration, while very improbable, is just as probable as any other configuration of marbles in the jar.  Yet, it is a necessity that one of these improbable configurations is instantiated in the jar, regardless of whether through material or intelligent causation.  As such, the mere fact that the configuration is improbable says nothing about the causal agency that brought it about.

CSI gets around this problem by placing a non-uniform probability distribution on sets of marble configurations.  So, while the particular configurations are equally probable in themselves, the probability of a configuration coming from a particular set varies.  Thus, with a the right choice of  specification, it is now possible to exclude either a material or intelligent causal agent as being responsible for the marble configuration.

But, how do we choose the correct specification?  The problem is that there are many possible specifications for the marble configurations.  In fact, there is a specification for each possible probability distribution.  The information conferred by picking one of these specifications must be averaged out over all possible choices of specifications.  Unfortunately, this brings us right back to square one, and each marble configuration again regains an equal, though very low, probability, when averaged over all choices of specifications.

This is where the problem with a subjective specification comes into play.

We might say that the selection of specification is up to the group of scientists detecting intelligent design, so even though syntactically speaking one specification is as preferable as any other, there is an external mapping of value to each specification (i.e. semantics) held by the scientists that makes one specification more valuable than another.  Furthermore, this mapping is held by the vast majority of people, and they all agree the mapping is objective.

Unfortunately, while this may well be true, such a mapping begs the question if it meant to support a scientific claim, since the mapping is no longer a strictly scientific claim.  Science, at least of the hard science variety, decides matters of qualification using quantification.  It proposes a variety of models to fit observed data, and uses model fitting (i.e. linear regression) to determine which model is best.  In this way, it is possible to make objectively true statements about the physical world, which is at the level of mathematical veracity, given the premises are true.

However, in the case of CSI, the choice of specification is what determines the answer to whether a particular object is designed.  Two choices of specification can give two completely different answers. Thus, if the choice of specification is determined by some factor completely external to the marble configurations under consideration, then this may be equivalent to fitting the data to the model, since the specification chosen is part of the data being measured.  Fitting the data to the model does not count as science.

As a result, even though it is not necessarily the case that an externally determined specification is fitting the data to the model, it remains a live possibility.  As long as there is no objective way to eliminate this possibility, then by the principle of maximum entropy both the possibility of fitting the data to the model and fitting the model to the data must be given equal weighting.  If both possibilities are given equal weighting then ultimately the CSI metric cannot give a positive reading, inherently cannot discriminate between design and non-design, and therefore does not count as a scientific claim.

The only way the CSI metric can count as science is if there is an objective way to choose a specification.


  1. Probability distribution is really helpful in understanding the non functional concepts.Otherwise theoretically it is difficult to learn the function theory.
    Paired t-test

  2. This is one reason why observable functional specificity that identifies islands of function in config spaces of possible arrangements -- overwhelmingly non-functional -- become a key point. In the case of the jar, the island where marbles on the perimeter create the glyphs to spell out in English/roman text, a valid word or expression requiring 500 + bits, defines such an island. The accessible blind chance and mechanical necessity search across arbitrary configs, will be maximally unlikely to hit on the island. This, at intuitive level, is why, say, we routinely distinguish a screen with intelligent message from lucky noise -- we cannot infer reasonably to the implied degree of the latter on the gamut of accessible resources on the 10^57 atom, 10^17 s scale of our solar system. KF