PaperlayerView exact revision on arXiv

Source-linked AI summary

DuoSketch: How Pairs Navigate Challenges in AI-Supported Collaborative Design Ideation

Weiyan Shi, Darryl Lim, Geraldine Quek, Kenny Tsu Wei Choo

arXiv:2610.10249v1cs.HC

TL;DR

How pairs work with AI as a shared design concept develops remains underexplored. DuoSketch combines a shared canvas and transcript with a separate AI for each collaborator, and a qualitative study follows 12 pairs. The study identifies three challenges and describes bridging work in both directions between AI proposals and shared design.

  • Problem

    Less is understood about the difficulties pairs encounter as they develop a concept through shared work and individual AI exchanges, and how they respond to those difficulties.

    Source: §1 · p. 2

  • Method

    DuoSketch paired a shared canvas and live discussion transcript with a separate AI for each collaborator, and an exploratory qualitative study followed 12 pairs developing concepts.

    Source: §1 · p. 2

  • Results

    Pairs faced mismatches in proposal fit, missing shared meanings, and omitted agreed changes, navigating them through bridging work between AI and shared design.

    Source: §7 · p. 17

  • Takeaways & Limitations

    The observed practices inform support for joint judgement, joint expression, and communication of evolving design context.

    Source: §7 · p. 17

    The study used short, co-located sessions with 12 pairs, so longer projects and established professional teams may develop and communicate shared intentions differently.

    Source: §6 · p. 17

Abstract

from arXiv · show

Generative AI offers new resources for collaborative design ideation, yet how designers work with it as a shared concept develops remains underexplored. We developed DuoSketch, integrating a shared canvas and live transcripts with a separate AI for each member of a pair. An exploratory qualitative study with 12 pairs identified three challenges: unresolved fit between AI proposals and shared design purposes, shared meanings not expressed in AI responses, and agreed changes missing from later outputs. We found that pairs navigated these challenges through bridging work in two directions: AI -> Shared Design (reworking features, developing uses, and assessing fit with shared requirements) and Shared Design -> AI (supplying materials, explaining meanings, and communicating decisions). We discuss how future AI systems can support the joint exploration and articulation of design ideas.

1 Introduction

Collaborative design ideation develops through discussion and shared artefacts, while AI adds material whose relevance can extend beyond the exchange that produced it. DuoSketch and a study of 12 pairs examine the challenges and bridging work involved as a shared concept develops.

1 Introduction

  • The process-oriented question is how collaborators continue working with AI while combining ideas, exploring changed situations, and preparing a concept together.Prior work includes multimodal design tools, shared AI workspaces, and visual collaboration tools.Source: §1 · p. 2
  • DuoSketch combines a shared canvas and live discussion transcript with a separate AI for each collaborator.Pairs could submit shared materials to AI and add generated images to the canvas; 12 pairs took part in an exploratory qualitative study.Source: §1 · p. 2
  • Pairs encounter unresolved fit with shared purposes, missing shared meanings in AI responses, and agreed changes absent from later outputs.The study examines how pairs navigate these challenges as their shared design develops.Source: §1 · p. 2
  • Pairs bridge AI and shared design by developing proposals within the design and communicating shared intentions back to AI.The account motivates support for joint judgement, joint expression, and communication of evolving design context.Source: §1 · p. 2

2 Related Work

Prior work connects generative tools to evolving design concepts, AI to shared work, and representations to joint design judgement. DuoSketch examines how pairs connect separate AI exchanges with shared discussion and sketching as their concept develops.

2.1 Generative Tools for Exploring and Developing Design Concepts

  • Concept-development tools support exploration through sketches, multimodal interaction, and iterative generation [45] [2].These approaches make artefacts and speech available as ways to express intentions and develop concepts.Source: §2.1 · p. 2
  • Modality preferences do not simply follow sketching’s benefits for exploration, and visual inspiration can increase fixation and reduce later concept diversity [52] [51].Generative material can both help express an intention and redirect it.Source: §2.1 · p. 3

2.2 AI Support for Shared Design Work

  • Group ideation and visual collaboration systems use shared workspaces and AI-supported artefacts to help people discuss concepts and communicate across differences in interpretation [41].These systems make shared activity a focal point for encountering and discussing AI contributions.Source: §2.2 · p. 3
  • Research on separate AI access highlights context drift between LLM instances, while multimodal design also involves shared sketches, speech, and generated references [20].DuoSketch examines how pairs connect separate AIs with shared discussion and canvas work.Source: §2.2 · p. 3

2.3 Shared Representations and Joint Design Judgement

  • Shared visual information does not itself establish agreement about significance; grounding, awareness, and knowledge translation offer lenses on joint understanding.Speech and sketching work together in communicating and developing design concepts.Source: §2.3 · p. 3
  • The analysis connects joint responses to AI proposals with the expression of shared intentions in requests, linking design development to communication.It treats these activities as bridging work across AI and shared design.Source: §2.3 · p. 3

3 Method

The study used DuoSketch with 12 pairs to examine how separate AI interactions relate to shared design development. It followed participants through staged ideation activities and analysed session records and interviews using challenge, navigation, and support codes.

3 Method

  • An exploratory qualitative study in Singapore followed 12 pairs developing design concepts with separate AIs and a shared canvas and live transcript.Shared design refers to concepts, requirements, and intended uses articulated and taken up through agreement, joint modification, or discussion.Source: §3 · p. 4

3.1 Participants

  • The 24 participants had design-course experience; 20 were aged 18–24, nine pairs knew each other, and three were formed for the study.Participants came from design, engineering, and humanities backgrounds and had no prior DuoSketch experience.Source: §3.1 · p. 4

3.2 DuoSketch: Shared Materials, Separate AI Chatbots

  • DuoSketch paired a jointly editable canvas and shared transcript with separate chatbots, allowing participants to send shared materials to AI and return generated images to the canvas.The two AI histories remained separate; selected canvas objects were sent as a snapshot, while transcript attachments included accumulated text.Source: §3.2 · p. 4

3.3 User Study

  • Pairs designed a portable product for everyday commuting through synthesis, adaptation to an added requirement, and consolidation after individual exploration.The study included a joint post-session interview revisiting concept progression, AI use, transfers, and design decisions.Source: §3.3 · pp. 4, 6

3.4 Qualitative Data Analysis

  • Researchers coded 36 pair–phase records across Phases 2–4, connecting challenges and navigation actions to discussion, sketches, and AI requests.Three challenge categories and six navigation strategies were developed and independently coded; interview suggestions received three support codes.Source: §3.4 · pp. 6–7

4 Results

The section examines how pairs navigated gaps between AI outputs and shared design, including how they developed and assessed proposals and made shared context, meanings, and decisions available to AI. It also covers their requests for support in jointly contributing to AI requests and communicating evolving design decisions.

4.1 Challenges Pairs Encounter in Connecting Shared Design with AI

  • Thirteen records across eight pairs showed C2, the most widespread challenge; overall, 20 of 36 pair–phase records across ten pairs received a challenge code.P8’s Phase 4 received both C2 and C3, yielding 21 challenge assignments; 16 records received no challenge code.Source: §4.1 · p. 9, §4 · p. 9, Table 3 · p. 11

4.1.1 C1: Unresolved Fit between AI Proposals and Shared Design

  • C1 arose in six records across five pairs when proposals reinterpreted a concept’s purpose, left a feature’s use unresolved, or introduced trade-offs against shared requirements.It occurred during Team Synthesis and Adaptation as pairs combined initial ideas or adapted concepts to new commuting situations.Source: §4.1.1 · p. 9

4.1.2 C2: Shared Design Meanings Not Expressed in AI Responses.

  • C2 affected eight pairs across thirteen records when AI responses omitted or misrepresented shared meanings, including component relationships, operating rules, and intended interactions.This could happen even when pairs supplied sketches or recognisable components, because requests or materials did not convey how those elements related to the design.Source: §4.1.2 · p. 9

4.1.3 C3: Agreed Design Changes Missing from Later AI Responses.

  • C3 arose for P8 and P11 during Consolidation when later AI outputs omitted an agreed combination or restored a rejected feature.P11’s request lacked the combined concept and its live transcript; P8’s updated sketch omitted solar panels, but the AI restored them until P8 explicitly stated the removal decision.Source: §4.1.3 · pp. 9, 11

4.2 How Pairs Bridge Shared Design and AI Interaction

  • Pairs bridged design and AI in two directions: they revised, developed, and assessed AI proposals, and supplied materials, explained meanings, or communicated decisions to AI.Five pairs used AI → Shared Design strategies, nine used Shared Design → AI strategies, and four pairs used both; the latter included reworking proposals, developing uses, assessing requirements, sharing relevant references, clarifying relationships, and instructing changes.Source: §4.2 · p. 12

4.3 Support Pairs Sought for Shared Design with AI

  • Three pair interviews sought support for jointly assessing or reopening proposals, including probing purpose and limitations, evaluating concerns with evidence, and revisiting earlier decisions with their AI context.These suggestions addressed the need to keep room and grounds for shared judgement as proposals became more concrete.Source: §4.3 · p. 14
  • Three pair interviews sought ways to participate directly in conveying shared meaning, beyond simply seeing the same materials or having context available to AI.Suggestions included letting both collaborators mark intended changes on an image and having AI ask questions when expression gaps arise.Source: §4.3 · pp. 14–15
  • Seven pair interviews raised selective access to AI context, including team chat alongside personal chats, sharing part of a history, and identifying who generated a response.Pairs also wanted to revisit earlier directions and share summaries or rationales, indicating that the source and relevance of transferred material mattered.Source: §4.3 · p. 15

5 Design Implications

The design implications translate observed challenges into opportunities to support joint judgement, expression of shared meanings, and communication of current design decisions. These opportunities are grounded in an exploratory study, and their effects remain to be evaluated.

5.1 Supporting Joint Judgement as Design Purposes Evolve

  • Five pairs encountered proposals whose fit with shared purposes was unresolved, and their responses included revising features, developing uses, and assessing requirements.Synthesis and Adaptation made proposal relevance salient, while DuoSketch lacked dedicated support for comparing proposals with purposes or recording unresolved trade-offs.Source: §5.1 · p. 16
  • Systems could help pairs assess proposals against evolving purposes by attaching intended uses, open questions, or criteria and comparing alternatives together.Earlier alternatives and the reasons for retaining or rejecting them could remain accessible, while evidence about cost or feasibility is linked to its source.Source: §5.1 · p. 16

5.2 Supporting Joint Expression of Shared Meanings

  • Eight pairs encountered responses that missed shared meanings, including a case where the AI received a diagram but not the coverage rule established in discussion.Separate chat panels meant shared canvas material and the meaning needed to interpret it could follow different paths into the request.Source: §5.2 · p. 16
  • Pairs could use a shared request draft to annotate visual references, explain relationships, and state intended changes before submission.Clarification questions could surface unspecified relationships, while participation remains flexible and collaborators retain separate exploration when desired.Source: §5.2 · p. 16

5.3 Supporting Communication of Current Design Decisions

  • Two C3 encounters involved agreed changes absent from later outputs, while seven pairs suggested selective access to context across AIs.The records identify different input conditions but do not establish which model processes produced the mismatches; evaluation should assess fidelity to current decisions and record-keeping costs.Source: §5.3 · p. 17
  • Systems could link decisions to sketches, discussion, and AI exchanges, then preview which version and decisions will accompany a request.Decision status, contributors, unresolved differences, and earlier conflicting statements could remain visible for pair review and correction.Source: §5.3 · p. 17

6 Limitations

The study’s scope and analytic approach constrain how its findings should be interpreted.

6 Limitations

  • The study involved 12 pairs in short, co-located sessions using DuoSketch; longer projects and established professional teams may develop and communicate shared intentions differently.All participants had design-course experience, and the structured synthesis, adaptation, and consolidation activities shaped the reported phase distributions.Source: §6 · p. 17
  • Challenge and navigation counts indicate coverage across pairs and phases, not request-level failure rates or strategy effectiveness.The categories and navigation matrix are interpretations of the sessions, C3 rests on two encounters, and focusing on challenges may underrepresent routine or preventive work.Source: §6 · p. 17

7 Conclusion

The study of 12 pairs identifies three challenges in AI-supported collaborative design ideation and describes how pairs bridged AI and shared design. These practices inform support for joint judgement, expression, and communication of evolving design context.

7 Conclusion

  • Pairs questioned the fit between AI proposals and shared purposes, clarified shared meanings missing from responses, and communicated agreed changes absent from later outputs.Source: §7 · p. 17
  • Pairs bridged AI and shared design by reworking, developing uses for, and assessing proposals, and by bringing shared materials, meanings, and decisions to AI.Source: §7 · p. 17
  • These practices inform support for joint judgement, joint expression, and communication of evolving design context.Source: §7 · p. 17

A A Four-Phase Design Example

Figure 13 provides a visual reference for P7’s canvas across the four study activities, showing sketches and generated images as the pair developed and consolidated a concept.

A A Four-Phase Design Example

  • The snapshots serve as a visual reference for the procedure, showing the pair’s concept development and consolidation.Source: A A Four-Phase Design Example · p. 20
  • Figure 13 presents P7’s canvas across Individual Exploration, Team Synthesis, Team Adaptation, and Team Consolidation.Source: Figure 13 · p. 20
  • The panels retain sketches, generated images, blue outlines, and handwritten annotations visible in the source figure.Source: Figure 13 · p. 20

B Data Corpus

The data corpus comprises records from 12 paired sessions and immediate post-session interviews. Reconstructed transcripts cover both the design sessions and interviews.

B Data Corpus

  • 12 pairs’ session materials and immediate post-session interviews form the study corpus.The records document participant and AI activity; interviews capture participants’ explanations and suggestions.Source: Table 6 · p. 21, B Data Corpus · p. 21

C Detailed Qualitative Codebooks

The codebooks distinguish challenges and responses observed during design activities from support suggestions made in interviews. They define three challenge codes, six navigation strategies, and three perceived-support codes.

C Detailed Qualitative Codebooks

  • C codes describe observed challenges, N codes observed responses, and S codes interview suggestions.C1–C3 concern shared purposes, meanings, and carrying agreed changes forward; N1–N3 develop AI proposals within shared design, while N4–N6 communicate shared design to AI.Source: C Detailed Qualitative Codebooks · p. 22, Table 7 · p. 22, Table 8 · p. 22
  • A storage compartment that would defeat a collapsible design’s purpose supports both C1 and N3; proposing collapsible and non-collapsible variants supports N1.N3 requires an articulated assessment against a shared purpose or constraint.Source: C Detailed Qualitative Codebooks · p. 22
  • A vague objection to a silicone strap supports C1 but not N3, while developing a cap-and-slot alternative supports N1.The distinction is between identifying difficulty and assessing an addition against an articulated shared requirement.Source: C Detailed Qualitative Codebooks · p. 22
  • Sending a newly developed screen-proximity condition to AI follows N2, while transferring an image and clarifying rack placement illustrate N4 then N5.The first example received no additional N4–N6 assignment because it addressed no previously identified expression mismatch; Table 4 counts challenge-linked actions.Source: C Detailed Qualitative Codebooks · p. 23
  • Perceived-support codes may overlap when interview suggestions contain distinct support needs.The codebook defines three perceived-support codes and provides interview examples.Source: Table 9 · p. 23
Loading 2610.10249v1…