II. The Strawman to Avoid
Before proceeding, we must explicitly reject an attractive but flawed argument.
The Strawman
Naur’s pessimism about documentation was valid for his time because keeping theory and code synchronized through documentation was a gargantuan undertaking that only humans could perform. Humans lack the time, energy, and cognitive capacity to maintain such documentation effectively. AI collaboration changes this equation by reducing labor costs, making theory preservation possible in ways Naur couldn’t have envisioned.
This argument misreads Naur. His pessimism is not rooted in labor costs but in the nature of tacit knowledge. If the problem were merely that documentation is laborious, then more effort (or AI assistance) would solve it. But Naur’s claim is stronger: certain knowledge cannot in principle be articulated, regardless of how much time or assistance one has.
The Impossibility Witnesses
Michael Polanyi, whom Naur cites, makes this vivid:
“It is pathetic to watch the endless efforts—equipped with microscopy and chemistry, with mathematics and electronics—to reproduce a single violin of the kind the half-literate Stradivarius turned out as a matter of routine more than 200 years ago.”
No amount of documentation labor could capture what Stradivarius knew tacitly.
Christopher Alexander reached the same conclusion in architecture. His Quality Without a Name—the property that makes spaces feel alive rather than dead—is immediately recognizable yet indefinable. Alexander spent years codifying patterns (A Pattern Language, 1977), producing 253 of them, yet acknowledged that following the patterns doesn’t produce the quality. As he wrote:
“There is a central quality which is the root criterion of life and spirit in a man, a town, a building, or a wilderness. This quality is objective and precise, but it cannot be named.”
The Convergence
Three domains—craft (Polanyi), architecture (Alexander), programming (Naur)—converge on the same epistemological wall. This is not coincidence but evidence of a structural limitation in formal systems.
Naur would likely respond to claims about AI: “AI is just another formal system; it cannot capture what humans themselves cannot articulate.”
We must not claim that AI “finally solves” the documentation problem. That would be to misunderstand what the problem is.