Web AI: From Generated Interfaces to Outcome-Informed Runtime AI Design

Artificial intelligence is increasingly capable of adapting, optimizing, personalizing, and generating user interfaces at runtime. These capabilities span established research traditions including adaptive interfaces, online experimentation, contextual bandits, reinforcement-learning-based interface adaptation, and Generative UI. This paper proposes Web AI as a framework for examining a further shift: from AI that primarily selects, adapts, or generates interface artifacts toward AI systems that learn from the measurable consequences of previous design decisions and use accumulated evidence to make subsequent design decisions at runtime for the current user context. The framework distinguishes runtime design decision-making from interface generation and introduces two partially independent capability dimensions: Learning Authority, the degree to which observed design consequences can alter future design decisions, and Design Authority, the degree to which AI possesses bounded operational control over what experience should exist. Mature Web AI is characterized by the conjunction of high Learning Authority and high Design Authority rather than by generative capability alone. Formally, runtime design is represented as D_t=F(K_t,X_t,G), where K_t is a persistent Learned Design State, X_t is the current User Context State, and G represents human-defined objectives, constraints, capabilities, and governance. The framework's central empirical signature is Delta KrightarrowDelta DrightarrowDelta Y: changes in accumulated design knowledge should be associated with changes in subsequent design decisions and, through those decisions, measurable outcomes. The contribution is a framework-level conceptual integration and research agenda rather than a claim that its underlying technical mechanisms are individually new. Keywords: Web AI; Adaptive User Interfaces; Generative UI; Runtime Design; Artificial Intelligence; Human–Computer Interaction; User Context; Reinforcement Learning; Contextual Bandits; Learning Authority; Design Authority; Outcome-Informed Design