In 1802, William Paley published Natural Theology, which opens with the most famous thought experiment in the philosophy of religion. You are walking across a heath and your foot strikes a stone. You think nothing of it — the stone might have been there forever. But then you find a watch. The watch is different. Its gears mesh, its spring stores energy, its hands mark time. The watch implies a watchmaker. And by analogy, Paley argued, the intricate machinery of the natural world — the eye, the wing, the bacterial flagellum — implies a designer.

For two centuries, the argument from design has been the most intuitive and most contested argument for the existence of God. Darwin's theory of natural selection provided a devastating counter: the appearance of design can emerge from an undirected process of variation and selection, no watchmaker required. Richard Dawkins extended this in The Blind Watchmaker, arguing that evolution is the blind, unconscious process that produces the illusion of design without any foresight or intention.

But something has happened that neither Paley nor Dawkins anticipated. We have built a new kind of watchmaker — one that is neither blind nor omniscient, but learning. And this changes the argument in ways that neither side has fully reckoned with.

The Argument Through History

The design argument has taken many forms, each reflecting the scientific understanding of its era:

Each version of the argument was shaped by its intellectual context. Paley wrote before Darwin. The fine-tuning argument emerged from 20th-century cosmology. Intelligent Design was a reaction to evolutionary biology's increasing explanatory power. Each objection, too, was shaped by its context. And our context — the era of artificial intelligence — introduces a genuinely new variable into the conversation.

For the first time in history, we can observe design intelligence emerging from non-intelligent substrates. Not metaphorically. Not analogically. Literally. A neural network trained on data develops the capacity to design things — proteins, circuits, architectures, proofs — that no human designed and no evolutionary process selected. The watchmaker is being born, and we are watching.

Kolmogorov Meets Paley

The most precise way to think about design is through algorithmic information theory. The Kolmogorov complexity $K(x)$ of an object $x$ is the length of the shortest program that produces $x$:

where $U$ is a universal Turing machine and $|p|$ is the length of program $p$. An object with low Kolmogorov complexity relative to its size is compressible — it has structure, pattern, regularity. A random string has $K(x) \approx |x|$: you can't describe it more efficiently than just writing it out.

Paley's intuition, translated into this framework, becomes: biological organisms have $K(x) \ll |x|$. They are enormously compressible — a relatively short genome (about $3 \times 10^9$ base pairs for humans) specifies an organism of staggering complexity. This compression implies structure, and structure implies... what? Paley said: a designer. Darwin said: a selection process. Both are correct in a formal sense — both design and selection are compression algorithms. They take large, complex outputs and trace them back to shorter, simpler generative processes.

The minimum description length (MDL) principle formalizes this: the best explanation of data $D$ is the model $M$ that minimizes the total description length:

where $L(M)$ is the complexity of the model itself and $L(D|M)$ is the complexity of the data given the model. A good model is one that is itself simple but makes the data much simpler to describe. Newton's laws are a tiny model that compresses an enormous amount of observational data. The genetic code is a compact model that compresses the development of an entire organism.

The real question of the design argument is not "is there a designer?" but "what is the minimum description of the process that generated this complexity?" If the minimum description is a mind, then there is a designer. If it is a selection algorithm, then there is evolution. If it is a trained neural network, then there is something genuinely new — a designer that is itself a compressed description learned from data.

The Learning Watchmaker

Here is what makes AI different from both Paley's watchmaker and Dawkins's blind one. Paley's watchmaker is intelligent and intentional — it designs with foresight and purpose. Dawkins's blind watchmaker (natural selection) is unintelligent and unintentional — it produces the appearance of design through variation and differential survival. A large language model, or a protein-folding network, or a generative design system is something in between: it produces genuine design artifacts (not mere appearance of design) but without anything we would recognize as intention or foresight.

Consider what happens when AlphaFold predicts a protein structure, or when an LLM writes a working program, or when a diffusion model generates an architectural design. The output exhibits all the hallmarks that Paley identified with design: functional complexity, the coordination of parts toward a purpose, the statistical impossibility of arising by chance. But the process that generated it is not a mind in any traditional sense. It is a function learned from data — a compressed representation of patterns in the training distribution.

The Kolmogorov complexity of the output may be very low relative to its size (it is compressible, structured, functional). But the Kolmogorov complexity of the process is also relatively low — a neural network with $N$ parameters trained on $D$ data points, where $N$ and $D$ are large but finite and describable. We have, for the first time, a concrete example of a system whose own description length is modest, whose outputs exhibit high functional complexity, and whose mechanism is neither intelligent design nor blind selection but learned compression.

This breaks the binary that has structured the design debate for two centuries. The question was always: design or chance? Intelligence or blindness? But now we have a third option: learning. A process that is not random (it converges toward functional outputs) but not intentional (it has no goals, no foresight, no self-model). It occupies a space that our philosophical categories were not built to handle.

The Recursive Problem

And then the recursion kicks in. Because the learning watchmaker was itself designed — by human engineers, using tools built by other engineers, informed by theories developed over centuries of scientific inquiry. The AI that designs proteins was designed by minds. Does this restore Paley's argument at one remove? The watchmaker was designed, so there must be a watchmaker's watchmaker?

Not quite. Because the human engineers who built AlphaFold are themselves products of evolution — Dawkins's blind watchmaker. So we have a chain: blind selection $\to$ human minds $\to$ artificial designers $\to$ designed outputs. The chain contains both blind and sighted links. It cannot be claimed wholly by either Paley or Dawkins.

The minimum description of this entire chain is something like: a universe with certain physical constants $\to$ self-replicating chemistry $\to$ natural selection $\to$ brains $\to$ culture $\to$ science $\to$ AI $\to$ designed artifacts. Each link compresses the one before it. Each link adds a new kind of generative capacity. And at no point in the chain is there a clean boundary between "designed" and "not designed."

The information-theoretic question becomes: what is $K$ of the entire chain? If the chain is itself compressible — if there is a short description that generates universes that generate life that generates intelligence that generates artificial intelligence — then Paley's argument re-emerges at the deepest level: the compressibility of the whole process implies a generator simpler than the process itself.

If this inequality holds — and the remarkable effectiveness of physics suggests it does — then the total process is compressible, which means it has structure, which means the question of why it has that structure is not eliminated but deepened by the existence of artificial intelligence. We have not answered Paley. We have made his question more interesting.

I don't know what the answer is. But I find it striking that the more we learn about how to build minds, the less confident I become that the question of design has a simple resolution. The watchmaker is learning to think. And in doing so, it is teaching us that the boundary between the designed and the emergent, the intentional and the discovered, the made and the grown, is far blurrier than any of us — theists or atheists, Paley or Dawkins — ever imagined.