Dijkstra's 1977 Warning: Why Natural Language Programming Fails in the AI Era

Edsger W. Dijkstra's 1977 essay "On the foolishness of 'natural language programming'" argues that formal symbol systems are indispensable for precise computation. The article reviews mathematical history: Greek mathematics stagnated because it remained at the textual and diagrammatic level; Islamic algebra declined after a brief flirtation with formal symbols and a return to verbal exposition; modern civilization emerged only after Europe broke free from medieval scholasticism — an obsession with verbal precision that failed — thanks to deliberately designed formal notations by Viète, Descartes, Leibniz, and Boole.

Dijkstra emphasizes that formal texts derive their power from a small set of simple syntactic rules, which efficiently eliminate the meaningless expressions that are almost unavoidable in natural language. Far from being a burden, formal symbols are a privilege: they enable ordinary students to achieve feats that once required genius. He cites a 1977 technical report preface that boasted of discarding standard logical connectives "for clarity," illustrating a widespread misunderstanding of this point.

A thought experiment follows: if natural language were the sole input/output medium for information-processing devices, history would repeat itself. Computer science would largely become an obscure craft — the study of how to gradually construct a sufficiently rigorous formal system starting from natural language. Compressing that interface to a usable width would demand all human wisdom, and even millennia might not suffice.

The author then connects this 40‑year‑old critique to current trends. The rise of Code Agents and "Vibe Coding" appears to challenge the idea that natural language can replace formal programming languages. The core issue is natural language's inherent ambiguity, redundancy, and imprecision. Can AI tools truly bridge the gap and resolve the conflicts that arise?

The article draws a distinction: both assembly/machine instructions and high‑level programming languages are formal languages; a compiler performs a semantics‑preserving, dimension‑reducing translation. The gap between a high‑level formal language and natural language, however, is fundamental. Having a large language model translate natural language into a high‑level language raises the question of whether semantic equivalence can ever be achieved — a question Dijkstra's analysis already answers.

A Zhihu answer by Wang Bo adds a vivid illustration: "If Dijkstra were alive today and saw us first using natural language to make AI write code, then investing enormous effort in building test, constraint, and verification loops to ensure the AI doesn't make mistakes, he would probably say: 'You finally built a machine that understands natural language, only to discover you still need formal methods to constrain it — isn't that exactly what I said?'"