International Research Journal of Engineering and Technology (IRJET)
e-ISSN: 2395-0056
Volume: 13 Issue: 02 | Feb 2026
p-ISSN: 2395-0072
www.irjet.net
Bidirectional Human AI Alignment for Generative User Interfaces and Design Developer Tooling Nadim Akhtar1, Laraib Ahmad Siddiqui2 1Assistant Professor, Computer Science Engineering, Chandigarh University 2MLOps Engineer, Accenture
---------------------------------------------------------------------***---------------------------------------------------------------------
Abstract - Generative user interfaces promise substantial
policies. In principle, Generative User Interfaces (GenUI) could compress the time between an early concept and a working prototype, enabling teams to explore more alternatives and to test ideas earlier with users.
gains in speed and breadth of exploration by allowing teams to produce interface concepts, component trees, and production code from natural-language intent. Yet current GenUI workflows are often brittle: they overfit to underspecified prompts, fail to respect product constraints such as accessibility and design-system rules, and provide limited visibility into why specific layout or interaction choices were made. These gaps are not merely model errors; they reflect misalignment between what people mean, what the system infers, and what downstream teams can validate and maintain. This paper frames the problem as bidirectional human AI alignment, in which the tool not only adapts to users but also helps users externalize intent as explicit, checkable commitments. We propose a mixed-initiative GenUI pipeline that converts multi-modal intent (text, screenshots, sketches, and existing UI code) into constraint artifacts, generates diverse UI variants linked to editable code, and performs verification against constraints including component policies, interaction invariants, and accessibility targets. Rather than relying on opaque prompting, the system surfaces traceable rationales, uncertainty markers, and targeted self-critiques so that designers and developers can negotiate trade-offs by adjusting constraints, not rewriting prompts. We outline an evaluation plan spanning controlled comparisons with prompt-only baselines, expert reviews with UX practitioners and front-end engineers, and longer deployments in real tool chains to study trust calibration, collaboration impacts, and maintainability. The goal is a GenUI workflow that is faster, more accountable, and easier to integrate into production design-to-code processes.
In practice, prompt-based UI generation can raise a familiar risk: speed rises while accountability falls. Misalignment often shows up in small choices that accumulate into real cost-a layout that ignores spacing tokens, a component choice that violates interaction policy, or a flow that looks plausible but fails keyboard navigation. These problems are not always visible at first glance. Many are discovered only when engineers integrate generated code into an existing codebase, when QA runs accessibility audits, or when design reviews compare outputs against a system of reusable components. The practical failure mode is therefore not that generation is impossible; it is that generation is easy to produce but hard to verify, explain, and maintain. The central question becomes not only “Can a model generate a UI?” but “Can the output be interrogated, corrected, and handed off responsibly?” Human-centered AI work emphasizes that trustworthy systems require legibility and user agency, not merely high average performance [1].
1.INTRODUCTION
Alignment is also socio-technical. UI work is distributed across roles, and each role interprets correctness differently. A designer may prioritize coherence, hierarchy, and brand consistency; an engineer may prioritize component reuse, testability, and future refactoring; a product manager may emphasize compliance and measurable outcomes; and an accessibility specialist may focus on semantics, focus order, and contrast. Prompt-only GenUI is typically optimized around one person’s immediate request, expressed in ambiguous language and without the surrounding organizational context that makes an interface “right.” Even when the result is acceptable, limited traceability makes it difficult to know which parts are safe to accept and which deserve scrutiny, producing “hidden rework” when issues surface downstream.
Generative systems have moved from novelty to daily infrastructure in many software teams, supporting ideation, prototyping, copywriting, documentation, and code generation. User interface work is a natural next frontier because UI artifacts are structurally rich and tightly constrained: interfaces must express product intent while obeying platform conventions, design-system standards, accessibility requirements, and a growing set of governance
Several research strands suggest a more robust direction. Mixed-initiative interface principles argue that effective systems share control: automation proposes actions while people guide, override, and refine behavior [3]. Guidelines for human–AI interaction emphasize showing system status and confidence, enabling efficient correction, and maintaining user control through iterative refinement [2]. Interactive machine learning likewise highlights that reliable
Key Words: Generative user interfaces, bidirectional alignment, mixed-initiative systems, design-to-code, constraint-based design, verification, explainable AI, developer tooling
© 2026, IRJET
|
Impact Factor value: 8.315
|
ISO 9001:2008 Certified Journal
|
Page 463