Aesthetic Taste and Its Limits: Breakdowns in Prompt-Mediated Design of User Interfaces
Abstract
How is taste expressed, and where does it live? When designers use AI tools like Lovable to generate interfaces from natural language, that question stops being abstract. Suddenly, taste has to fit inside a sentence.
Our Human-Centered Design study investigates how user experience (UX) designers express aesthetic taste when using AI-powered interface generation tools. Through thirteen observed design sessions, we identify six recurring breakdown types that characterize how aesthetic intent fails during prompt-mediated generation: Vocabulary Gap, Execution Gap, Convergence Trap, Tacit Ceiling, Authorship Disconnect, and Risk-Reward Freeze. Our analysis further introduces four designer profiles—Reference-Led, Specification-Led, Efficiency-Led, and Process-Invested—that predict susceptibility to these breakdowns based on early-session prompting strategies. We find that the gap between aesthetic judgment and linguistic expression is a structural condition of prompt-mediated design, where authorship depends more on the process of making than final output quality. To address these tensions, we propose four directions and a maturity assessment for future generative UI tools: supporting visual input as primary specification, enabling local and intent-anchored editing, promoting safer iteration through versioning, and making the co-creative process and authorship visible.
The Aesthetics of Prompting
How is taste expressed, and where does it live? When designers use AI tools to generate interfaces from natural language, that question stops being abstract. Suddenly, aesthetic taste has to fit inside a sentence.
Our project investigates the challenges UX designers face when translating complex, tacit visual judgment into linear, explicit prompts. We specifically examine the "repair moment," the critical site of human-AI collaboration where designers must diagnose visual inadequacy in a generated output and decide whether to correct it through re-prompting or manual editing.
We argue that the struggle to express taste through generative UI tools is not a matter of poor prompting, but a fundamental tension in prompt-mediated interaction.
- Taste is hard to express verbally. Designers often know what they want before they can say it, and prompt-based tools force that judgment into language too early.
- Polished is not the same as personal. AI-generated interfaces can look finished while still feeling generic, borrowed, or unlike the designer's own work.
Authorship lives in the process of creation. A design that matches the brief can still fail to feel authored when the act of making is delegated to the system.
Key Findings: introducing a new taxonomy
Through thirteen observed design sessions, we found that this translation process breaks down in six recurring ways, compromising creative control and undermining authorship:
- Vocabulary Gap: Designers lack the precise language for their intent.
- Execution Gap: The tool fails to deliver on a clear, well-phrased prompt.
- Convergence Trap: The system defaults to generic, "polished" styles.
- Tacit Ceiling: Aesthetic judgment reaches a level that resists any verbalization.
- Authorship Disconnect: The designer feels removed from the act of making.
- Risk-Reward Freeze: Designers avoid ambitious prompts due to high risk of failure
These breakdowns are experienced differently depending on a designer's initial approach. We introduce four Designer Profiles that correlate with susceptibility to these failures.
- Reference-Led: Relies on external examples for style.
- Specification-Led: Focuses on detailed component lists.
- Efficiency-Led: Prioritizes speed and minimal iteration.
Process-Invested: Prioritizes feeling in control of the creative process.
Directions for Future Generative UI Tools
To address these structural conditions and better support genuine aesthetic expression, we propose a new tool and four actionable directions for tool development:
- Integrated Designer Taste Maturity Tool to assess and guide the development of aesthetic taste expression in various dimensions.
- Support visual input as primary specification (reducing reliance on language).
- Enable local and intent-anchored editing (improving the repair moment).
- Promote safer iteration through versioning (mitigating Risk-Reward Freeze).
- Make the co-creative process and authorship visible (closing the Authorship Disconnect).