Generative User Interfaces in AI-Driven Design
- Generative User Interfaces are digital interfaces created through AI-driven processes that transform abstract inputs like prompts, sketches, or models into dynamic interactive designs.
- They employ diverse methodologies—from LLM-driven structured interactions to GAN-based layout synthesis—to adapt interfaces responsively to complex user tasks and evolving contexts.
- Research indicates that generative UIs accelerate rapid prototyping and iterative design while also posing challenges in consistency, human control, and safety oversight.
Generative User Interfaces are interfaces that are created, refined, or evolved through computational generative processes, with AI systems contributing to design, implementation, or adaptation. In current research, the term covers several closely related phenomena: LLM-produced task-specific interfaces that replace linear chat with structured interaction, high-fidelity UI mock-ups generated from high-level descriptions, model-driven interfaces synthesized from formal specifications, and malleable interfaces that continue to evolve as users revise goals, data, and context. Across these formulations, the central move is from manually authoring a fixed interface toward generating screens, components, interaction flows, code, or executable UI actions from abstract intent such as prompts, sketches, screenshots, domain models, or semantic specifications (Lee, 21 May 2025, Chen et al., 26 Aug 2025, Chen et al., 22 Jan 2025).
1. Conceptual foundations
The contemporary literature frames Generative User Interfaces as a distinct HCI paradigm rather than a narrow code-generation technique. One line of work defines a Generative User Interface as a user interface “created, refined, or evolved through computational generative processes,” emphasizing collaborative creation among human designers, AI assist tools and agents, and final users. Another defines it operationally as an LLM-produced UI tailored to the user’s task, where the response is an interactive, structured interface rather than a static text answer. A third treats GenUI as malleable software: systems that adapt to diverse needs by producing interfaces on demand rather than shipping a single fixed design (Lee, 21 May 2025, Chen et al., 26 Aug 2025, Min et al., 25 Jan 2026).
Several recurring tensions organize the field. The first is the tension between expressiveness and discoverability. Prompt-only generation preserves breadth but obscures what can be customized, while menu-based customization becomes unwieldy when “almost everything” can vary. The second is the tension between generative breadth and human control: users struggle to articulate multidimensional design intent through linear prompts, then struggle again to interpret why a generated interface looks and behaves as it does. This is described as a gulf of execution and a gulf of evaluation, with semantic drift arising when iterative refinements produce disproportionate and contradictory changes. The third is a task-structure tension. Linear request–response chat is repeatedly described as inefficient for multi-turn, information-dense, exploratory, and stateful tasks, which motivates interfaces that externalize state, expose controls, and support manipulation directly (Park et al., 27 Jan 2026, Chen et al., 26 Aug 2025, Min et al., 25 Jan 2026).
A further conceptual distinction separates GenUI from adjacent ideas. It is not identical to classical adaptive UI, because recent work places equal emphasis on design-time co-creation and runtime adaptation. It is not identical to code generation, because several systems generate screenshots, layouts, semantic models, or declarative UI protocols rather than source code alone. It is also not identical to retrieval-based inspiration systems, although such systems can participate in a generative workflow. This suggests that “Generative User Interfaces” is best understood as an umbrella term for systems that use generative models to synthesize the interface artifact itself, the interaction substrate that surrounds it, or both (Lee, 21 May 2025, Zhao et al., 2021, Mozaffari et al., 2022).
2. Historical trajectory and research lineages
The longer historical backdrop places GenUI within broader UI evolution. One survey traces a progression from rule-based and text interfaces in the 1960s and 1970s, to the GUI era of the 1980s and 1990s, touch-first mobile UIs in the 2000s, voice-based assistants in the 2010s, and multimodal, AI-augmented interfaces in the 2020s. Within this trajectory, recent GenUI research marks a shift from pre-authored interaction structures toward interfaces that are generated, reorganized, or adapted at runtime from multimodal inputs and context (Bieniek et al., 2024).
Early generative strands were already present in model-driven UI engineering. GENIUS treated UI generation as a transformation pipeline from combined Task/Domain/Abstract interaction descriptions into a running interface, with usability criteria embedded directly in the transformation rules and refined a posteriori through expert and user evaluation. Its significance lies less in neural generation than in the idea that transformations themselves are design artifacts carrying ergonomic intent, and that a generated interface can be governed by explicit intermediate models, state charts, and runtime interpretation (Sottet et al., 2013).
Later work introduced data-driven visual generation and component recomposition. GUIGAN moved away from pixel synthesis toward component-level generation, encoding a screen as a sequence of reusable GUI component subtrees and combining adversarial learning with style compatibility and structure validity signals. GANSpiration, by contrast, used a style-based GAN to support design inspiration rather than full interface synthesis, balancing targeted and serendipitous example generation through layerwise style mixing, clustering, and optional nearest-neighbor retrieval of real UI screenshots. These systems established two important research directions: generation as composition over structured UI parts, and generation as a support for exploratory design work rather than only final artifact production (Zhao et al., 2021, Mozaffari et al., 2022).
By 2024–2026, the literature diversified rapidly. Some systems centered on intent-based outcome specification from sketches, prompts, and themes; some focused on interactive, generated interfaces for LLM-assisted tasks; some emphasized gradual exposure of customization layers; some treated explicit semantics or task-driven data models as the substrate of generation; and some targeted personal agents that emit lightweight, executable UI actions. This diversification indicates that GenUI is not converging on a single canonical architecture. Instead, it is branching into multiple lineages: model-driven generation, neural layout or code generation, mixed-initiative design tools, generated task interfaces, and high-stakes domain-specific interfaces (Zhang et al., 2024, Chen et al., 26 Aug 2025, Cao et al., 6 Mar 2025, Kong et al., 24 May 2026).
3. Architectures and intermediate representations
A defining feature of GenUI systems is their reliance on intermediate representations that mediate between intent and rendered interface. These representations vary substantially. In generated task interfaces for LLMs, the interface is specified as an interaction graph , where nodes are views or subgoals and edges are transitions, together with per-component finite state machines and a structured JSON schema for metadata, states, elements, events, and flows. In malleable information-task interfaces, the central object is a task-driven data model , where is an object-relational schema, a dependency graph, and structured data conforming to them. In semantic-guidance systems, the intermediate representation is a hierarchical semantic structure spanning Product, Design System, Feature, and Component levels, linked by a relationship graph with match, conflict, and missing-value edges. In gradual-generation systems, the representation is decomposed into stage-specific specifications: natural language for categories, JSON schemas for layout, code for content logic, and CSS variables for style (Chen et al., 26 Aug 2025, Cao et al., 6 Mar 2025, Park et al., 27 Jan 2026, Min et al., 25 Jan 2026).
Other systems choose different scaffolds. Frontend Diffusion inserts a Product Requirements Document between sketch and code, using a three-stage pipeline of sketching, writing, and coding, with iterative refinement over generated code. GUIGAN represents a GUI as a depth-first sequence of component subtrees augmented by structure strings and learned style embeddings. Macaron-A2UI adopts a declarative action protocol in which the model emits a single JSON object containing text_response and a stream of structured UI actions such as beginRendering, surfaceUpdate, dataModelUpdate, and deleteSurface. GENIUS relies on a model-driven chain from TDA to State-Chart and CUI to FUI. These choices differ in formalism, but all externalize some portion of the latent generation process into a manipulable, inspectable substrate (Zhang et al., 2024, Zhao et al., 2021, Kong et al., 24 May 2026, Sottet et al., 2013).
This plurality matters because many of the field’s core problems are representational rather than merely generative. Prompt-only systems obscure customization structure; semantically guided systems seek to make intent explicit and outcomes interpretable; task-driven data models seek continuity across revisions; gradual-generation systems seek discoverability by exposing layers; protocol-based personal-agent systems seek renderability and safety by constraining output grammar. A plausible implication is that progress in GenUI depends not only on more capable models, but also on better intermediate representations that can be validated, rewound, compared, and localized under change (Park et al., 27 Jan 2026, Min et al., 25 Jan 2026, Cao et al., 6 Mar 2025).
4. Interaction patterns and design workflows
Survey work on generative AI applications organizes user-guided interaction into prompting, selection, system or parameter manipulation, and object manipulation. These primitives combine into higher-level UI layouts such as conversational interfaces, canvases, contextual interfaces, modular or pipeline interfaces, and simulated or spatial interfaces. In conversational layouts, history supports iterative refinement; in canvas layouts, direct manipulation localizes edits; in modular layouts, blocks and chains externalize flow; in contextual layouts, assistance appears in place; and in simulated layouts, spatial gestures and embodied interaction become first-class inputs (Luera et al., 2024).
Empirical studies of design practice show that GenUI tools are often used in two distinct workflows. One begins by generating multiple screens up front and then refining them; the other generates one screen at a time through incremental dialogue. Participants in the GenUI Study found the first mode useful for breadth but often overwhelming, with irrelevant features and poor fit to precise intent, while the second gave greater control. The same study reports that GenUI’s current “sweet spot” is the rapid production of a “good enough” first draft, but that last-mile editing, cross-screen consistency, design-system adherence, and integration with incumbent tools such as Figma or IDEs remain major barriers (Chen et al., 22 Jan 2025).
User-centered design research on “vibe coding” situates GenUI inside an AI-in-the-loop ideate–prototyping process. In that case study, low-fidelity wireframes and design workshops were followed by LLM-driven generation of React-based prototypes using V0 and Bolt.new. Interactive prototypes elicited richer, more actionable feedback than static sketches, and the team used them as design probes rather than production artifacts. This work positions GenUI not as a replacement for UCD, but as a way to compress the path from intent to testable interface while keeping human orchestration, debugging, and domain grounding central (Li et al., 28 Jul 2025).
Several papers generalize these patterns beyond design tools. Application-style educational interfaces replace free-form chat with structured inputs, prompt templates, validation views, traffic-light triage, and inline editing to manage complex, high-stakes feedback workflows. Automotive attentive interfaces propose subtle, context-adaptive microinterventions—interactive scenarios, conversational primers, personalized takeover requests, subtle nudges, and ambient scene generation—to improve Situation Awareness in Level 3 driving without overt workload-increasing alerts. AR interfaces for generative design infer constraints from direct manipulation in situ and project feed-forward previews of the resulting constraint state. Personal-agent interfaces synthesize lightweight, executable UI surfaces for slot collection, preference refinement, confirmation, and multi-goal organization. Together these systems broaden GenUI from mock-up generation to the generation of interaction surfaces embedded inside larger decision and control loops (Pozdniakov et al., 2024, Ebel, 2024, Kang et al., 27 Mar 2025, Kong et al., 24 May 2026).
5. Empirical evidence and application domains
Evidence for GenUI is heterogeneous: some papers report controlled evaluations with quantitative results, while others remain conceptual or formative. In artifact generation, GUIGAN reports that it significantly outperforms its best baseline by 30.77% in Frechet Inception distance and 12.35% in 1-Nearest Neighbor Accuracy, with category-scenario averages of FID 0.075 versus 0.113 and 1-NNA 0.869 versus 0.980. In a pilot user study, it also outperformed FaceOff on aesthetics, harmony, structure, and functionality, with average functionality 0.812 versus 0.452 (Zhao et al., 2021).
In LLM-mediated task interfaces, “Generative Interfaces for LLMs” reports strong human preference over conversational baselines across 100 queries. GenUI wins 84% overall against a Claude 3.7 conversational baseline, 69% against GPT-4o, and 75% against an instructed-UI baseline, with especially strong results in Data Analysis & Visualization at 93.8% and Business Strategy & Operations at 87.5%. The same study attributes much of this advantage to cognitive offloading and perceived professionalism, and reports that moving from one-shot generation to iterative refinement improves overall human win rate by +14% (Chen et al., 26 Aug 2025).
In controllability and refinement, the semantic-guidance study reports statistically significant improvements over a chat baseline across intent expression, output interpretation, and ease of modification. For example, Intent Expressiveness rises from to with , Output Interpretability from to 0 with 1, and Ease of Modification from 2 to 3 with 4. These results support the claim that explicit semantic representations can bridge the gulfs of execution and evaluation in AI-driven UI generation (Park et al., 27 Jan 2026).
In malleable task interfaces, the Jelly system reports a technical evaluation over 25 task scenarios, with 197 entities, 1052 attributes, and 232 dependencies. For less detailed prompts, 94.12% of entities and 93.91% of attributes were rated “necessary and expected”; dependency relationships were 89.17% correct and mechanisms 98.33% correct. A user study with 5 found that all participants agreed or strongly agreed that the generated information was relevant, that customization was easy, and that layout and organization were effective (Cao et al., 6 Mar 2025).
In personalization, the field has documented substantial divergence in UI preference judgments. The personalization study collected 12,000 pairwise judgments from 20 trained designers and found mean pairwise agreement 0.624 with Cohen’s 6 for binary preferences, and lower agreement for four-way labels. Its adaptive designer-mixture model achieved 0.620 accuracy at onboarding budget 7 and, in an online generation study with 12 new designers, a 60.35% aggregate win rate over profile-conditioned and zero-shot baselines (Peng et al., 10 Apr 2026).
Application-specific evidence is likewise growing. Feedback Copilot, an application-style educational interface, reports that an advanced prompting and validation pipeline significantly improved average feedback quality over a base version, with 8, 9, and means of 0 versus 1; the strongest gains were in empathetic and self-reflective feedback dimensions (Pozdniakov et al., 2024). Macaron-A2UI reports that its best model reaches 75.6 overall on A2UI-Bench without explicit schema hints, surpassing a full-schema frontier baseline (Kong et al., 24 May 2026). Other systems remain earlier-stage: Frontend Diffusion explicitly reports no quantitative measurements or formal user studies, and gradual-generation work presents prototypes and future evaluation questions rather than participant data (Zhang et al., 2024, Min et al., 25 Jan 2026).
6. Limitations, governance, and future directions
The literature is unusually explicit about risk. Safety-critical domains raise reliability, latency, and certification issues. Automotive takeover interfaces must gate generative outputs with verified sensor facts, bound response times, avoid hallucinations, and manage false positives and false negatives. Personal-agent UI protocols enforce schema validity, renderability, and confirmatory controls because malformed or over-complex UI actions directly affect transaction flow. Educational applications emphasize pseudonymization, human approval, provenance inspection, and scope limitation away from grading. These cases treat GenUI not as unconstrained content generation, but as interaction generation under governance constraints (Ebel, 2024, Kong et al., 24 May 2026, Pozdniakov et al., 2024).
Across domains, recurring limitations include hallucinations, bias, prompt brittleness, reproducibility problems, context-window limits, maintainability of generated code, and weak support for highly branded or novel styles. The GenUI Study reports failure modes in intent assimilation, cross-screen consistency, editing, and originality. The vibe-coding case study highlights cumulative errors in long sessions and the need for session resets, debugging, and code review. Semantic-guidance work notes confusion around category boundaries and the compression of nuanced intent into fixed slots. Personalization work shows that even trained designers use shared concepts such as hierarchy or cleanliness differently, which makes fixed rubrics insufficient for individual preference modeling (Chen et al., 22 Jan 2025, Li et al., 28 Jul 2025, Park et al., 27 Jan 2026, Peng et al., 10 Apr 2026).
Open research questions cluster around standardization, selective use, and deeper integration. One line seeks standardized UI DSLs and portable structured representations for flows, states, events, and bindings. Another asks when a query merits a generative interface rather than chat, to avoid unnecessary complexity and latency. Others focus on richer backend integration, more reliable multi-frame design-to-code transformation, empirical comparison of AI-in-the-loop design workflows against traditional UCD, and the definition of common intermediate layers across application types. High-stakes domains add questions about evaluation of non-deterministic systems, regulatory compliance, and the validation of subtle or unconscious interventions (Chen et al., 26 Aug 2025, Li et al., 28 Jul 2025, Min et al., 25 Jan 2026, Ebel, 2024).
A broad pattern nevertheless emerges. GenUI research increasingly treats the interface not as a fixed endpoint, but as a generated, revisable, and inspectable artifact embedded within human-AI collaboration. Systems differ on whether they generate code, layouts, semantic models, executable UI actions, or subtle contextual interventions; they also differ on whether they prioritize creativity support, workflow acceleration, personalization, or operational safety. Yet the field converges on a common principle: useful generative interfaces do not merely produce outputs. They externalize intent, expose controllable structure, preserve provenance, and support iterative human judgment at the point where generation meets action (Lee, 21 May 2025, Cao et al., 6 Mar 2025, Chen et al., 26 Aug 2025).
