FeedQUAC: Quick Unobtrusive AI-Generated Commentary

TL;DR

FeedQUAC tackles the challenge of constant design feedback by delivering ambient, real-time AI commentary through multiple personas read aloud while designers work. The authors implement a lightweight, floating duck UI that cycles feedback automatically or on demand, using GPT-4 Vision and ElevenLabs voices to provide context-aware, diverse perspectives. A design-probe with eight experienced 3D CAD designers shows the system offers inspiration, validation, and efficiency, while highlighting issues around design-stage context and user control. The study discusses how ambient feedback can complement human critique and outlines design and evaluation considerations for future ambient creativity-support tools. Overall, the work advances ambient CST design by demonstrating feasibility, user acceptance, and the nuanced trade-offs between interference and support in creative workflows.

Abstract

Design thrives on feedback. However, gathering constant feedback throughout the design process can be labor-intensive and disruptive. We explore how AI can bridge this gap by providing effortless, ambient feedback. We introduce FeedQUAC, a design companion that delivers real-time AI-generated commentary from a variety of perspectives through different personas. A design probe study with eight participants highlights how designers can leverage quick yet ambient AI feedback to enhance their creative workflows. Participants highlight benefits such as convenience, playfulness, confidence boost, and inspiration from this lightweight feedback agent, while suggesting additional features, like chat interaction and context curation. We discuss the role of AI feedback, its strengths and limitations, and how to integrate it into existing design workflows while balancing user involvement. Our findings also suggest that ambient interaction is a valuable consideration for both the design and evaluation of future creativity support systems.

Figures (11)

• Figure 1: Framework based on the analysis of design forums, illustrating the components involved in (Q)uestion posts for feedback requests (left, orange) and (A)nswer posts for feedback responses (right, yellow). Question posts typically include visual context (sketches, images, and links), textual context (design motivations, goals, and background), and various types of questions (ranging from project-specific to career-oriented). Answer posts, responses, usually offer supportive messages (validation and encouragement), critiques and solutions addressing the posted questions, resources (worked examples and tutorials), and suggestions tailored to the project or career goals.
• Figure 2: Annotation FeedQUAC User Interface Walkthrough: Tanya, a 23-year-old entry-level 3D CAD designer, is designing her first egg chair in Fusion 360. Unsure about the curvature of the chair's bottom base, she missed her TA’s office hours and finds online feedback too slow. She turns to FeedQUAC for real-time guidance. To start, she opens FeedQUAC and accesses the control panel for setup. She selects the TA-like constructive Mentor persona for feedback and chooses to capture the entire active window of Fusion 360 , avoiding distractions from other open applications. For frequency, she sets emoji feedback every 30 seconds and audio feedback every 3 minutes . She also ensures the system retains the last two feedback generation as part of the context to avoid repetition . As she works, a set of egg and chair emojis appears from the duck icon, humorously acknowledging her design . Curious, she manually triggers feedback using Command + R. The duck icon at the bottom right rotates, and after 5 seconds, a supportive female voice encourages her progress: "It’s encouraging to see your 3D modeling skills applied... Considering refining the base..." The spoken feedback also appears as text next to the icon. Encouraged, Tanya adjusts the chair’s bottom base shape. Three minutes later, without manual pressing the hotkeys, the duck automatically rotates again, delivering another round of mentor-like supportive feedback.
• Figure 3: An illustration of the eight personas used in FeedQUAC, along with their icons and corresponding keywords used in the personality prompts. The three personas on the leftmost orange column cover different levels of positivity: constructive Mentor, supportive Cheerleader, and grumpy Critic. The three personas in the middle yellow column focus on different aspects of design: aesthetic Designer, practical Analyst, and innovative CEO. The two on the rightmost green column represent different humanness levels: sassy Friend and robot-like No Persona (AI). See the full gpt-4-vision-preview prompts and ElevenLabs voice IDs and descriptions in Appendix A.1 and A.2. See examples of each persona's feedback in Appendix A.3.
• Figure 4: Visualization of feedback requests for each participant during the design sessions, showing the selected persona, the timing of feedback requests, and the mode of generation---either automatic (within the three-minute feedback cycle, represented as circles) or $\blacksquare$ manual (via Command + R, represented as squares) . The timeline for each participant is aligned with their first feedback request for consistency in comparison. On average, participants received 13.25 feedback instances per session, with 5.26 instances manually requested and the remaining 7.99 instances generated automatically. Some participants, such as P2 and P4, relied solely on the automatic feedback cycle, while others, like P6 and P8, requested manual feedback via hotkeys for more than half of their feedback requests. These participants also manually explored different personas after reviewing the early feedback. Among all persona requests, Critic (31 times) was the most frequently used persona.The most frequently manually requested personas were No Persona (AI) (12 times) and Critic (11 times).
• Figure 5: Distribution of Likert scale responses for participants' overall user experience using FeedQUAC (a), their perceived level of pressure and stakes when gathering feedback using the AI system (b), and when receiving traditional human feedback (c). For readability, "Neutral" responses are omitted, and unselected options are blurred in the legend.