PaperlayerView exact revision on arXiv

Source-linked AI summary

Design Principles for Generative AI Applications

Justin D. Weisz, Jessica He, Michael Muller, Gabriela Hoefer, Rachel Miles, Werner Geyer

arXiv:2401.14484v1cs.HCcs.AI

TL;DR

The paper addresses the need for guidance that helps people interact with generative AI applications effectively and safely. It introduces six design principles developed iteratively through conceptual analysis and empirical work, finding that they supported actionable improvements across generative AI applications.

  • Problem

    The paper addresses the need for general design guidelines that help people interact with generative AI applications in effective and safe ways.

  • Method

    The paper introduces six design principles and develops them iteratively through critical conceptual analyses and empirical work.

  • Results

    The principles helped practitioners generate useful and actionable design improvements across generative AI applications producing text, images, and music.

  • Takeaways & Limitations

    The principles provide design guidance applicable to a range of generative AI applications.

  • Takeaways & Limitations

    The modified heuristic evaluation focused on commercially available generative AI applications.

Abstract

from arXiv · show

Generative AI applications present unique design challenges. As generative AI technologies are increasingly being incorporated into mainstream applications, there is an urgent need for guidance on how to design user experiences that foster effective and safe use. We present six principles for the design of generative AI applications that address unique characteristics of generative AI UX and offer new interpretations and extensions of known issues in the design of AI applications. Each principle is coupled with a set of design strategies for implementing that principle via UX capabilities or through the design process. The principles and strategies were developed through an iterative process involving literature review, feedback from design practitioners, validation against real-world generative AI applications, and incorporation into the design process of two generative AI applications. We anticipate the principles to usefully inform the design of generative AI applications by driving actionable design recommendations.

1 INTRODUCTION

Generative AI introduces a new interaction paradigm in which users specify desired outcomes while models control how computation is performed. The paper responds with six design principles, practical strategies, and validation intended to support effective and safe generative AI applications.

  • Generative AI lets users specify desired outputs while models determine how computation is performed.
  • Existing HCI guidelines have not addressed the specific nuances of generative AI.
  • The paper introduces six principles for designing generative AI applications, including new considerations and reinterpretations of known AI issues.
  • Each principle includes practical strategies and examples implemented through design processes or specific UX capabilities.
  • The principles were validated through multiple rounds of testing, feedback collection, iteration, and use in two generative AI applications.

2 RELATED WORK

HCI has developed broad, technology-specific, and AI-focused design guidelines across successive interaction paradigms. However, existing AI guidance has limitations in comprehensiveness and operationalization, while generative AI remains insufficiently addressed.

  • Guidelines span general UX, AI-infused systems, specialized technologies, application domains, user populations, and ethical concerns.
  • HCI guidelines evolved alongside computing paradigms including terminals, graphical interfaces, the Web, mobile systems, and AI.
  • AI guidelines can improve user experiences and address ethical challenges, but studies critique their comprehensiveness and operationalizability.
  • Existing HCI AI guidelines primarily focus on discriminative AI and do not account for generative systems that produce artifacts as outputs.

3 WHY GENERATIVE AI NEEDS DESIGN PRINCIPLES

Generative AI needs dedicated design principles because it combines outcome specification, variable outputs, new interaction skills, and risks that existing guidelines do not fully cover. The paper seeks principles applicable across generative AI domains and technologies.

  • Generative AI uses intent-based outcome specification: users describe desired results without specifying how they should be produced.
  • Generative variability means outputs may differ in character or quality even when the user’s input remains unchanged.
  • Prompt engineering is an emerging skill typically developed informally through trial and error.
  • Open-ended prompts and variable outputs create challenges for achieving desired and replicable results.
  • Generative AI introduces risks including intellectual-property concerns, harmful content, sensitive-information disclosure, malicious code, and underrepresentation.
  • The paper proposes general principles to help practitioners design safer and more effective applications across generative AI domains and technologies.

4 DESIGN PRINCIPLES FOR GENERATIVE AI APPLICATIONS

The paper presents six high-level design principles for generative AI UX, combining new interpretations of existing AI concerns with principles addressing generative-specific issues. Each principle is paired with strategies that can guide design processes or UX features.

  • Each principle is coupled with four strategies implemented through design processes or specific features and functionality.
  • The six principles include Design Responsibly, Design for Mental Models, and Design for Appropriate Trust & Reliance.
  • The generative-specific principles are Design for Generative Variability, Design for Co-Creation, and Design for Imperfection.
  • The principles and strategies support optimization goals focused on task-specific criteria and exploration goals focused on inspiration and alternative possibilities.
  • Design practitioners are expected to use judgment about whether a principle and its strategies apply to a particular use case.

5 METHODOLOGY

The paper develops design principles for generative AI applications through an iterative process intended to provide designers with specialized language, actionable strategies, and risk awareness.

  • The framework aims to provide designers with language for generative-AI-specific UX issues.
  • It offers concrete strategies and examples for difficult design decisions involving model capabilities and user needs.
  • The framework sensitizes designers to risks and processes for avoiding or mitigating harms.
  • The principles were developed through literature review, feedback, heuristic evaluation, and application to two generative AI applications.
  • Across iterations, the authors discussed prior feedback and revised the principles and strategies through organizational or wording changes.

6 ITERATION 1: CRAFTING INITIAL DESIGN PRINCIPLES

The first iteration combined literature on generative AI, design guidelines, and human-AI interaction with commercial application analysis to identify relevant design characteristics and organize them into principles and strategies.

  • The authors searched research literature and workshops on generative AI, design guidelines, human-centered AI, co-creation, explainability, and creative interfaces.
  • They also examined commercial generative applications to identify common design patterns and connect user needs with supporting UX designs.
  • The authors introduced a two-tier structure in which principles state important characteristics or considerations and strategies specify how to implement them in UX.
  • The analysis identified generative-AI characteristics involving multiple outputs, imperfect outputs, user control, and exploration of possibilities.
  • Existing AI concerns such as participatory design, explainability, and the AI’s role in co-creation were treated as especially relevant to generative AI.
  • The initial framework contained 7 high-level principles and 22 strategies, while allowing strategies to overlap across principles.

7 ITERATION 2: EXTERNAL AND INTERNAL FEEDBACK

External workshop and internal practitioner feedback drove revisions to the framework, including task-specific principles, reframing explainability, and separating process-oriented strategies.

  • Feedback from an approximately 50-person workshop and an internal guide viewed by over 1,000 practitioners informed the second iteration.
  • The revisions recognized that users’ goals can differ and added two task-specific principles.
  • The authors distinguished Design for Exploration for ideation, exploration, and learning from Design for Optimization when a singular artifact is desired.
  • Explainability was incorporated into Design for Appropriate Trust & Reliance rather than treated as an end in itself.
  • The authors rephrased strategies as action rules and identified five strategies concerning the design process rather than specific UX capabilities.
  • At the end of Iteration 2, the framework contained 8 high-level principles and 29 specific strategies.

8 ITERATION 3: MODIFIED HEURISTIC EVALUATION

A modified heuristic evaluation assessed the framework’s clarity, relevance, and coverage across commercial generative AI applications using design-practitioner evaluations.

  • 8.1 Method: The evaluation targeted clarity for design practitioners, relevance to commercial applications, and gaps in the framework.
  • 8.1 Method: The study examined 9 commercial generative AI applications selected for popularity, supported modalities, and use as either core experiences or embedded components.
  • 8.1 Method: 18 design practitioners with varied roles and experience evaluated applications individually and remotely using a modified heuristic exercise.
  • 8.1 Method: Evaluators identified examples before seeing the provided strategy labels, then labeled examples and rated relevance and clarity.
  • 8.2 Results: The evaluation produced 18 canvases containing real-world examples and notes about difficulty or confusion.
  • 8.2 Results: 11.9 examples were found for each strategy on average, and every strategy had at least one example, suggesting relevance across applications.
  • 8.2 Results: Evaluators generally rated the principles relevant and clear, while identifying 16 label mismatches, eight overlap issues, and three new strategies.
  • 8.2 Results: Five evaluators reported overlap between most strategies in Design for Exploration and Design for Optimization and strategies in other principles.

9 ITERATION 4: APPLICATION TO GENERATIVE AI UX DESIGN

Structured workshops assessed whether six design principles and accompanying strategies could support generative AI UX design in practice. Participants generated actionable ideas across both early ideation and later evaluation contexts, while feedback identified needs for richer guidance and user research.

  • Method: The principles and strategies were evaluated through two structured workshops with design practitioners working on generative AI applications.One team worked on a later-stage prompt-testing environment, while another designed an internal LLM-based conversational tool in an earlier ideation phase.
  • Workshop outcomes: The principles and strategies generated varied ideas across all six principles and produced concrete proposals for improving generative AI products.Examples included prompt effects that bake in content and systems for recognizing and sharing strong prompts.
  • Actionability: Practitioners described the resulting ideas as actionable, including potential roadmap items, newly identified blind spots, and quickly generated requirements.These comments connect the workshops to practical product planning and risk identification.
  • Design-process implications: Participants said applying the principles would be easier with more explanatory resources, examples from existing tools, broader team involvement, and user research.They linked user research to understanding users’ expectations for model outputs and identifying concrete design ideas.
  • Actionability: The principles could be applied in both early ideation and later evaluation stages to drive actionable design ideas.Workshop outcomes supported application across different phases of design.

10 DISCUSSION

The paper develops six generative AI design principles and 24 implementation strategies through iterative conceptual and empirical work. The principles supported actionable improvements across diverse generative AI applications, while the authors identify overlap, evaluation, and lifecycle boundaries for future refinement.

  • Contributions: The authors identify six principles with 24 companion strategies for designing generative AI applications.They developed them iteratively through critical conceptual analysis and empirical work aimed at scientific validity and real-world utility.
  • Results: The principles helped practitioners generate useful, actionable design improvements across applications generating text, images, and music.The authors report applicability across different media types and a range of generative AI applications.
  • Limitations and future work: The framework has boundaries: strategies can support multiple principles, heuristic evaluation sometimes confused principles with products, and the work examined only commercial applications.The authors also note that the guidance currently focuses on UX rather than model selection, tuning, deployment, monitoring, or policy decisions.
  • User goals versus design principles: The framework distinguishes optimization and exploration as different user goals, while recognizing that principles support both to varying degrees.Design for Imperfection is especially aligned with optimization, Generative Variability with exploration, and Appropriate Trust & Reliance with optimization in high-stakes domains.
  • Practical use: The principles and strategies function as a toolbox that practitioners can apply holistically or selectively when crafting generative AI experiences.The framework is intended to provide vocabulary for understanding and designing for new kinds of generative AI use.

11 CONCLUSION

The paper presents six design principles for generative AI applications, distinguishing reinterpretations of known AI issues from issues unique to generative AI. The principles pair with implementation strategies and were developed iteratively through literature, practitioner feedback, application validation, and design-process use.

  • Six principles address generative AI application design, with three reinterpreting known AI-system issues and three identifying generative AI-specific issues.
  • Each principle includes strategies implemented through specific UX features or a design process.
  • The principles and strategies were developed through literature review, practitioner feedback, validation against real-world applications, and use in designing two applications.
  • The framework addresses generative AI’s rapid incorporation into existing applications and new products, aiming to help practitioners harness it safely and effectively for users.

A EXTENDED DESCRIPTIONS AND EXAMPLES

The extended descriptions explain each principle and strategy for design practitioners, using realistic examples drawn from commercial and experimental generative AI systems. The authors also describe heuristic evaluation and note a boundary on process-related examples.

  • Extended descriptions and examples clarify the meaning and application of each design principle and strategy for design practitioners.
  • Examples come from commercial generative AI applications or experimental generative AI systems, including the listed commercial systems.
  • A modified heuristic evaluation supplied most examples, while later iterations added examples for three new strategies.
  • Process-related strategies were discussed theoretically because the authors lacked visibility into the actual design processes of examined applications.
  • A single UX feature or functionality may implement more than one design strategy, so similar examples may recur.

A.1 Design responsibly

The paper makes responsible design the most important principle because generative AI systems can produce diverse harms, especially for vulnerable people. It recommends a socio-technical perspective that evaluates technical mechanisms against user value.

  • Responsible design is identified as the most important principle for generative AI systems.
  • Generative AI may produce diverse harms, with particular concern for people in vulnerable situations.
  • A socio-technical perspective asks whether proposed technical mechanisms improve user experience, provide capabilities, or address user pain points.
  • Designers should avoid technosolutionism, the idea that technology will solve all human problems.

A.1.1 Use a human-centered approach*.

The section describes human-centered strategies for responsible generative AI design, including understanding stakeholders and mental models, managing emergent behaviors and harms, and supporting exploration of varied outputs.

  • Use a human-centered approach*: Human-centered approaches examine users’ workflows and pain points so proposed generative AI uses align with actual needs.
  • Use a human-centered approach*: Value Sensitive Design helps identify stakeholders and navigate value tensions among end users, builders, and purchasing or licensing decision-makers.
  • Use a human-centered approach*: Designers must balance narrowly defined capabilities against open-ended interaction that can surface risky emergent behaviors.
  • Use a human-centered approach*: Generative models can produce toxic, unfair, or inaccurate outputs, so designers should evaluate risks and provide reporting and escalation mechanisms.
  • Use a human-centered approach*: Useful mental models should explain system behavior, including varied and potentially non-reproducible outputs, while reflecting users’ backgrounds and goals.
  • Use a human-centered approach*: Tutorials, examples, explanations, and social transparency can help users learn to work effectively with generative AI applications.
  • Use a human-centered approach*: Evaluating users’ mental models can reveal misunderstandings, while existing interaction patterns can support learning.
  • Use a human-centered approach*: Capturing user expectations, behaviors, preferences, and background can support more personalized interactions through additional prompt information.
Loading 2401.14484v1…