TL;DR
Prototyping AI-powered interfaces is challenging because AI's dynamic nature makes traditional design tools insufficient. This paper introduces ProtoAI, a novel workflow and tool that integrates AI model exploration directly into the UI prototyping process, allowing designers to experience AI behavior firsthand and iteratively refine designs based on model outputs and potential failures.
What Problem Does This Paper Address?
Designing user interfaces (UIs) for AI-powered applications (AIX) presents unique challenges compared to traditional software. AI's behavior is inherently dynamic, influenced by training data, user interactions, and evolving models. This makes it difficult for designers to anticipate how an AI will perform under various conditions and how to best present its capabilities and limitations to end-users. Current UI prototyping tools often treat AI as a "black box," forcing designers to use separate tools to understand AI models, leading to a disconnect between interface design and AI behavior. This friction hinders rapid, iterative design and the creation of effective, human-centered AI experiences that can handle uncertainty, explain their actions, and recover from errors.
Methods
The paper proposes Model-Informed Prototyping (MIP), a workflow that bridges the gap between AI model exploration and UI prototyping. Their system, ProtoAI, implements this workflow through four main views:
- AI Models and Services View: Allows designers to select and configure available AI models or services.
- Data View: Enables importing diverse input data for model simulation and defining personas or scenarios. This view also supports data transformations and the creation of calculated columns.
- UI Designer View: A visual interface where designers construct the prototype. Crucially, it allows direct incorporation of AI model outputs (e.g., confidence scores, heatmaps) as UI elements.
- Data Previews View: This view automatically renders the UI prototype across various input data contexts, simulating how the interface will behave with different user inputs and model outputs.
ProtoAI's "design-by-instance" approach lets designers directly experience the AI's behavior as they design, rather than relying on static placeholders.
Datasets
The paper does not explicitly list specific datasets used for training or evaluation. However, it describes how designers can import "diverse input data for model simulation" and "portrait photos for each persona taken across different usage contexts." This suggests that the system is designed to work with user-provided datasets, which can include images, feature data, and potentially ground truth labels for AI models like face recognition.
Key Insights and Ideas
- Bridging the AI-Design Chasm: The core insight is that effective AIX design requires designers to directly interact with and understand the AI model's behavior, not just its final output. ProtoAI facilitates this by making AI model outputs tangible elements within the prototyping process.
- "Fail Fast, Fail Often" for AIX: The Data Previews view enables rapid evaluation of designs across numerous scenarios, allowing designers to quickly identify AI breakdowns and iterate on solutions, embodying the "fail fast, fail often" principle for AI features.
- Operationalizing Human-Centered AI (HAI): By integrating model behavior into the design process, ProtoAI helps designers operationalize HAI guidelines, such as providing explanations, handling failures, and managing user control, directly within their prototypes.
- Data as a Design Material: The system treats input data and model outputs as materials for design, allowing designers to explore how variations in data affect the UI and AI performance.
Key Results
The paper focuses on the workflow and system design rather than presenting quantitative results from user studies or benchmark evaluations. The key "results" are the capabilities of the ProtoAI system:
- Designers can directly incorporate AI model outputs into UI prototypes.
- The system automatically generates data previews of the UI across diverse input data contexts.
- Designers can evaluate interface choices against different user inputs and contexts of use.
- The system supports iterative revision of designs by analyzing model breakdowns and creating conditional interface states.
Limitations and Future Work
Limitations:
- The paper is an extended abstract, so a deep dive into empirical validation or quantitative results is not provided.
- The focus is on the workflow and tool; the complexity of integrating various AI models and the scalability of the system are not detailed.
Future Work:
- Investigating how AIX designers can drive AI model parameters based on UI features.
- Exploring negotiation of model outputs necessary for explainability.
- Improving communication of discovered failure instances between designers and engineers.
- Extending ProtoAI to support Responsible AI needs like fairness, accessibility, and transparency.
- Developing a library of widgets implementing AIX design patterns and explainable overlays to scaffold learning.
Summary updated