Why Architecture May Miss the AI Revolution: Consilience, Constraint-Governed Design, and Incorrigibility

Abstract

Objective: Architecture may fail to benefit from generative artificial intelligence unless AI operates on top of an evidence-tested design theory. Visual novelty is not architectural innovation. Methods: This paper uses Boyd’s OODA learning loop and synthesizes Alexander’s theory of living structure from The Nature of Order, Salingaros’s extensions of living geometry and pattern-language structure, evidence from environmental psychology and visual-attention studies, organizational-learning theory, and a documented AI-assisted museum design case. Findings: Alexander’s fifteen Fundamental Properties provide a geometrical tool for wholeness and coherence that can be made operational in LLM prompts; diagnoses using these criteria agree with emotion-based LLM judgments, public preference, and eye-tracking evidence. Constraint-governed AI can generate, compare, criticize, and revise designs against salutogenic criteria. This consilience contrasts with mainstream architectural approval systems, which remain siloed from experiential and scientific constraints and utilize ungoverned AI to accelerate stylistic recycling. Contributions: AI use in architecture is reframed as an epistemological problem. The New National Museum of Ecuador counterproposal is examined as a theory-first demonstration using pattern grammar, form grammar, experiential narrative, and iterative AI critique. Limitations: The museum remains a concept-stage demonstration requiring independent replication and pre-occupancy testing — concept-stage evaluation before occupancy, using VR methods, AI analysis, and eye tracking of design images. Applications: Pattern-based, auditable AI design pipelines, and evidence-linked procurement. Social impact: constraint-governed design intelligence provides a framework for making effects on human well-being explicit and empirically testable, whereas unconstrained AI use risks accelerating production while leaving the effects on users outside the learning loop.