CeleryKills

How a Borrowed Vocabulary from Information Theory Masks Circular Reasoning in Arguments for Intelligent Design

There is a familiar rotation in the arguments assembled against evolution and abiogenesis, a pattern that has repeated often enough over the last half century to feel almost regional, like a tide chart or a salmon run. First came the insistence that gaps in the fossil record reveal discontinuity rather than incompleteness. Then the claim that the second law of thermodynamics forbids increasing order in biological systems. A third move points to “irreducible complexity,” where biological structures are said to fail if any part is removed. A fourth pivots to improbability calculations, asserting that life’s origin is too unlikely to occur by chance. A fifth reframes evolutionary theory as “just a theory,” leaning on colloquial meaning rather than scientific use. Each of these, at various moments, has been presented not simply as critique but as disproof. I have watched these ideas circulate through classrooms and community discussions with a kind of persistence that is less about data than about framing.

I worked in information science and data system, ending with early AI and machine learning, so this next ploy is one I understand fairly well.

Into this lineage arrives “specified complexity,” a term introduced and developed primarily by William Dembski in the late 1990s and early 2000s. It is presented as a quantitative method to detect design, an attempt to formalize intuition about patterns that are both unlikely and meaningful. The idea feels engineered to resonate with technical audiences while remaining rhetorically flexible. That dual character is part of its traction. I am struck by how often it is invoked in discussions of DNA, as if the mathematical language itself confers empirical grounding.

Dembski defines specified complexity as the conjunction of low probability and an independently given pattern. If an event is both highly improbable and matches a “specification,” then, under his framework, design is inferred (Dembski, The Design Inference, 1998). The structure borrows vocabulary from information theory while modifying its logic. The problem appears immediately when one asks how probability is assigned. Probability relative to what distribution? Biological systems do not sample uniformly from all possible molecular configurations. They evolve through constrained pathways shaped by chemistry, selection, and history. Assigning a uniform distribution over an astronomically large space and then declaring any observed structure improbable is not an empirical calculation. It is a category choice that predetermines the outcome.

Consider a concrete evolutionary pathway rather than an abstract probability space. Early tetrapods transitioning from aquatic to terrestrial environments did not explore all conceivable anatomical configurations. They operated within narrow biochemical and developmental constraints. Limb structures, for instance, emerged through incremental modification of existing fin bones, producing variations that were viable in shallow-water or unstable shoreline habitats. Some lineages, like the ancestors of modern amphibians, persisted and diversified. Others, equally “possible” in a combinatorial sense, disappeared, leaving no descendants. The Devonian record is crowded with these partial experiments, organisms that occupied specific ecological niches and then vanished as conditions shifted or competitors emerged. The probabilities that matter here are not uniform draws from all anatomical permutations. They are conditional probabilities embedded in a sequence of contingent events. Each step depends on prior structure, environmental pressure, and reproductive viability. Many branches terminate. A few continue.

A similar pattern appears at the molecular level. Protein families evolve through duplication, mutation, and divergence. Enzyme functions shift gradually, often retaining partial activity before acquiring new specificity. Some variants fold improperly and are degraded. Others function poorly and confer no selective advantage, effectively disappearing from the population. A minority stabilize and propagate because they interact productively with existing cellular systems. These are not isolated rolls of a cosmic die. They are biased walks through a constrained landscape, where the distribution itself changes as the system evolves. Extinction is not an anomaly in this framework. It is an expected outcome of exploration under constraint. When viewed this way, the appeal to a single, global probability becomes incoherent. The system never samples that space to begin with.

The notion of “specification” compounds this issue. In formal contexts, specification would require an independent description that is not tailored after observing the data. Yet in practice, the patterns cited for DNA or proteins are derived from the observed sequences themselves. This reverses the direction of inference. One does not begin with a pattern and then test whether a sequence matches it; one observes the sequence and then constructs a pattern that fits. The independence condition is quietly abandoned, though it is essential to the logic (Elsberry and Shallit, “Information Theory, Evolutionary Computation, and Dembski’s ‘Complex Specified Information’,” 2011).

Examples from molecular biology show how “specification” is typically reconstructed after the fact. The canonical helix–turn–helix motif in DNA-binding proteins is identified by aligning known regulatory proteins and abstracting a consensus pattern from their shared structural features, not by starting with an independently defined motif and then searching blindly for matches. Zinc finger domains follow a similar path. Their repeating cysteine and histidine residues were recognized only after sequencing and structural studies revealed a recurring binding configuration; the “pattern” is distilled from observed instances. Even at the level of DNA, promoter regions such as the bacterial −10 and −35 consensus sequences are statistical summaries derived from collections of functioning promoters. They tolerate variation and are defined by frequency distributions, not fixed templates specified in advance. I find that these cases all move in the same direction. Researchers observe functional sequences, compare them, and then formalize the common elements as patterns. The specification emerges from the data. It does not precede it, which means it cannot serve as an independent criterion for detecting design.

To understand where the language comes from, it helps to revisit the foundations of information theory. Claude Shannon, working at Bell Labs in 1948, defined information as a measure of uncertainty in a message source. The Shannon information content, or entropy , of a discrete random variable with probabilities is given by:

(Shannon, “A Mathematical Theory of Communication,” 1948). This measure does not evaluate meaning or function. It quantifies the average uncertainty, which is why it is so effective in telecommunications and, more recently, in domains like fraud detection where irregularities in distributions signal anomalies. A transaction stream that deviates sharply from expected entropy profiles can indicate engineered interference rather than typical variation. I have seen how practitioners treat Shannon entropy as a diagnostic of distributional shifts, not as a measure of purpose.

A practical illustration comes from network traffic analysis. Suppose a monitoring system tracks the distribution of packet sizes or inter-arrival times across a large volume of normal activity. Under stable conditions, these features exhibit a characteristic probability distribution shaped by user behavior and protocol constraints. The Shannon entropy calculated over that distribution provides a baseline of uncertainty. When an intrusion or coordinated fraud attempt begins, it often imposes structure on what was previously diffuse. Packets may arrive in uniform bursts or with repeated signatures, reducing entropy in some dimensions while spiking it in others. Analysts do not interpret this shift as evidence of intent by reading meaning into the bit patterns. They compare the observed entropy against the expected distribution and flag statistically significant deviations. The signal is the divergence itself, not any intrinsic “purpose” embedded in the data.

Kolmogorov complexity, developed in the 1960s by Andrey Kolmogorov, Ray Solomonoff, and Gregory Chaitin, takes a different approach. It defines the information content of a string as the length of the shortest program that can produce it on a universal Turing machine. Formally, the Kolmogorov complexity of a string is:

where is the length of program and is a universal machine (Li and Vitányi, An Introduction to Kolmogorov Complexity, 2008). This concept underpins digital compression. When a file compresses efficiently, it contains regularities that allow a shorter description. When it resists compression, it is closer to algorithmic randomness. There is no appeal to meaning here either. A random string can have high Kolmogorov complexity without conveying function.

A straightforward example appears in everyday file compression utilities. Consider a large text document filled with repeated phrases or predictable structures, such as log files from a server. Compression algorithms like Lempel–Ziv scan the file and replace recurring substrings with shorter references, effectively constructing a compact program that reproduces the original data. The compressed file is shorter because its Kolmogorov complexity, approximated through the algorithm, is lower than the raw representation. By contrast, take a file of the same size composed of cryptographically secure random bits. It will not compress in any meaningful way, because there are no exploitable regularities. I have tested this myself with simple tools. The output size barely changes. The key point is that compressibility reflects the existence of internal structure that can be encoded more concisely, not the presence of semantic content or intended function.

“Specified complexity” attempts to combine these frameworks but does so by conflating their domains. It treats low probability in a hypothetical distribution as equivalent to high Shannon information, then overlays a notion of specification that resembles compressibility or pattern, loosely echoing Kolmogorov complexity. The synthesis is not formalized in a way that preserves the rigor of either theory. Instead, the hybrid becomes a rhetorical instrument. The calculation often reduces to asserting that a biological structure is both improbable and functional, then labeling that conjunction as design.

When this is applied to DNA, the misclassification becomes more evident. DNA sequences are not random strings drawn from a uniform distribution. They are products of iterative processes. Mutation introduces variation. Selection filters it. Genetic drift shifts frequencies. Constraints of biochemistry limit the accessible space. To ask for the probability of a particular sequence without modeling these processes is to step outside the system being studied. It is like calculating the probability of a specific river path by assuming water ignores gravity and terrain.

There is also a persistent confusion between information and function. Shannon information does not measure biological function. Kolmogorov complexity does not distinguish between meaningful and meaningless strings. DNA’s capacity to encode proteins is a result of mapping between nucleotide sequences and biochemical processes. That mapping is established through evolution. Invoking “information” without specifying which definition is being used creates a sliding scale where any structured outcome can be described as informational in a way that implies intention.

Historically, the emergence of specified complexity coincides with legal and educational challenges to teaching creationism in public schools in the United States. The rebranding to “intelligent design” sought to frame arguments in ostensibly scientific language. I notice how the adoption of mathematical terminology operates as a boundary marker. It signals rigor without always delivering it. The courts have recognized this shift as strategic rather than evidentiary (Kitzmiller v. Dover Area School District, 2005).

Philosophically, the argument leans on an intuition that complexity with purpose must arise from intention. That intuition is not derived from information theory. It precedes it. The mathematics is then used to give the intuition formal clothing. This is where theological assumptions enter. The inference to design presupposes an agent capable of producing the specified outcome. That agent is not part of the model, not constrained by its parameters, and not independently testable. The argument substitutes a conclusion for an explanation.

I find myself asking two questions that seem unavoidable. First, can a coherent, process-aware probability model for biological sequences be constructed that preserves the independence condition required by specification? Without that, the improbability claim lacks footing. Second, can any definition of information that is both mathematically rigorous and biologically relevant be shown to entail intentional design, rather than emergent function? If either question resolves negatively, the structure of specified complexity collapses. If either resolves positively, it would require a reformulation far more precise than what is currently offered.

There is a quiet irony here. The tools borrowed to support the argument, Shannon entropy and Kolmogorov complexity, are among the most successful abstractions in modern science because they refuse to encode meaning or intention. They measure structure and uncertainty without metaphysical commitments. When those tools are repurposed to assert design, they are asked to do something they were explicitly built to avoid.

One effective way to dismantle the argument in practice is to bring the discussion back to its required conditions and hold them there. The first condition is a well-defined probability distribution grounded in the actual generative process. If the distribution is hypothetical or uniform over an unconstrained space, the calculation is not about biology at all. The second condition is an independently specified pattern, articulated before the data are examined. If the “specification” is derived from the observed sequence, the inference collapses into circularity. I have found that simply restating these requirements often changes the tone of the exchange. The argument depends on both conditions, yet neither is typically satisfied when “specified complexity” is applied to DNA or proteins.

A second approach is to separate the borrowed mathematics from the conclusion being drawn. Shannon entropy measures uncertainty in a distribution. Kolmogorov complexity estimates compressibility. Neither one encodes purpose, agency, or intention. When the argument slides from “this system has structure” to “this system must be designed,” it introduces a premise that is not present in either framework. Asking where, exactly, that premise enters the calculation tends to expose the gap. Is it in the probability model? In the definition of specification? Or is it an external assumption that has been smuggled in under technical language?

It also helps to insist on consistency. If specified complexity reliably detects design, it should do so across domains where generative processes are well understood. Yet when applied to outputs of evolutionary algorithms, compression systems, or stochastic simulations, the method does not distinguish between designed and emergent structure without prior labeling. That inconsistency is difficult to reconcile with claims of objectivity. I notice that the argument often retreats at this point, shifting back to intuition about biological function rather than maintaining the formal criteria.

A concise statement that resolves the issue might read like this:

The concept of specified complexity fails as an inferential tool because it cannot simultaneously provide a process-grounded probability distribution and an independently defined specification; without both, its designation of “design” is not derived from the data but assumed in advance, rendering the conclusion unsupported and the argument moot.

That statement does not rely on counter-assertion. It points to the internal requirements of the method and shows that they are not met. Once that is clear, the rhetorical force of the term dissipates.

I think about how often the conversation returns to the same terrain under new names. The language shifts, the equations appear, the framing tightens. Yet the underlying move remains a substitution of assertion for demonstration. It is not an unfamiliar pattern.

References

Dembski, W. A., The Design Inference, 1998.
Elsberry, W., Shallit, J., “Information Theory, Evolutionary Computation, and Dembski’s ‘Complex Specified Information’,” 2011.
Shannon, C. E., “A Mathematical Theory of Communication,” 1948.
Li, M., Vitányi, P., An Introduction to Kolmogorov Complexity and Its Applications, 2008.
Kitzmiller v. Dover Area School District, 2005.
Chaitin, G. J., Algorithmic Information Theory, 1987.
Solomonoff, R. J., “A Formal Theory of Inductive Inference,” 1964.
Cover, T. M., Thomas, J. A., Elements of Information Theory, 2006.
Adami, C., “What Is Information?” Philosophical Transactions of the Royal Society A, 2016.
Schneider, T. D., “Information Content of Individual Genetic Sequences,” Journal of Theoretical Biology, 2000.
Altschul, S. F. et al., “Basic Local Alignment Search Tool,” Journal of Molecular Biology, 1990.
Lenski, R. E. et al., “The Evolutionary Origin of Complex Features,” Nature, 2003.