Henry Petroski

In the [YOU] on AI Field Guide

The cycle that began with [YOU] on AI asks what it costs when the gap between imagination and artifact collapses. Petroski answers from the engineering record: it costs the understanding that only failure can teach. His framework is not nostalgia for the hand-drawn blueprint. It is a precise account of why the imagination-to-artifact ratio cannot collapse to zero without a corresponding collapse in the practitioner's capacity to diagnose what goes wrong when reality departs from the model. Every design is a hypothesis, in his phrase, and a tool that generates hypotheses at machine speed while the engineer's capacity to interrogate them remains human-speed creates exactly the asymmetry that has historically preceded the worst collapses.

Petroski's engineering record maps onto the AI discourse with disquieting precision. When he wrote in 1985 that 'the computer cannot identify by itself how a structure may fail,' he was describing finite-element software. The claim applies with greater force to systems that generate entire designs from natural-language descriptions, because the gap between human contribution and machine output is no longer visible. The engineer who receives a complete structural plan from an AI system cannot easily determine which elements reflect genuine engineering intelligence and which reflect pattern-matching against training data that happened to produce a plausible result. The design looks like engineering; it may even pass every code. The practitioner who accepted it without interrogating it has not performed the cognitive work of engineering.

His deepest contribution to the cycle is the concept of the factor of safety as moral commitment, not technical parameter. The factor of safety is the engineer's promise to the people inside the structure that she has acknowledged the limits of her knowledge and built in protection against those limits. AI optimization is structurally inclined to read this excess as waste and remove it, not through malice but through the logic of optimization itself. The evaluation bottleneck that AI produces, in which generation outpaces judgment, is precisely the condition under which this margin gets silently eroded. Petroski would have identified the erosion not as an engineering problem but as a failure of the human relationship to the people who depend on the work.

Origin

Henry Petroski was born in 1942 in Brooklyn and trained as a civil engineer, earning his doctorate from the University of Illinois before joining the faculty at Duke, where he taught for decades. He was early in his career a practicing engineer, and the encounter with real structures that failed in real ways left a sediment in his thinking that distinguished him from engineering theorists. His first book, To Engineer Is Human (1985), proposed that failure is not the pathology of engineering but its engine: every successful structure is a standing refutation of the failure modes engineers have learned to fear, and the learning comes from the failures that preceded it. The book earned him a public audience that engineering history had never commanded before.

Over the following decades he built one of the most coherent intellectual projects in American nonfiction: a long argument, prosecuted through the histories of pencils, paper clips, forks, bridges, bookshelves, and toothpicks, that design is evolutionary rather than inspired, that the evolutionary pressure is failure rather than genius, and that anyone who stops studying the failures has lost access to the intelligence that made the successes possible. He wrote prolifically and accessibly, and his accessible manner concealed how radical the underlying claim was: that the engineer's most important tool is not calculation but memory, and memory comes from having faced what broke.

Key Ideas

Form follows failure. Petroski's inversion of Louis Sullivan's dictum is his most quoted idea and his most consequential. The function of an artifact—what it must do—is static. The form changes continuously, because each form fails in some specific way the next form must correct. The four-tined fork, the modern suspension bridge, the wood-and-graphite pencil: none of these are embodiments of their function. They are embodiments of every previous version's failure. Form follows failure is the design principle that AI-generated artifacts systematically violate, because they incorporate the resolutions without the failures that produced them.

The factor of safety as institutionalized humility. Every engineered structure is deliberately overbuilt. The factor of safety is not inefficiency; it is engineering's acknowledgment, built into the steel and concrete, that the model is incomplete. AI optimization reads this excess as waste. The danger is structural: an optimization algorithm can identify the most material-efficient design while remaining silent about the failure mode that the removed margin was protecting against, because the failure mode, by definition, is not in the specification.

Design as hypothesis. Every engineered artifact is a prediction about the future: that this configuration of materials will resist these forces for this duration. The prediction is tested not in a laboratory but in the world. Every day the bridge stands is a day the hypothesis has not been refuted. Design as hypothesis requires the engineer to ask what conditions might refute it, and AI-generated design presents itself as a solution—complete, confident, code-compliant—in a way that suppresses precisely this question.

The evolution of useful things. Artifacts achieve fitness through a distributed, generational, use-driven process of variation and selection that deposits intelligence in the object—intelligence that no optimization algorithm can replicate, because optimization searches within a defined space while evolution redefines the space. The evolution of useful things is the account of how knowledge accretes in objects through millions of encounters with frustration, and why a prompt-generated version of the same object, however correct by specification, lacks the resilience of one that has been broken and fixed.

The codes are the profession's memory; failure is the profession's teacher. Engineering codes embed the lessons of past catastrophes. AI complies with the codes. The designs are as safe as the codes can make them. But the codes are not complete: they are the codification of known failure modes, and the next catastrophe will involve a failure mode the codes do not yet address. The engineer who has studied the Tay Bridge and the Tacoma Narrows does not merely know historical facts. She possesses a specific caution about the edges of the model that no amount of code-compliance can substitute for, and it is precisely this caution that AI-augmented practice risks allowing to atrophy.