GraftMind: Facilitating Group Ideation with AI-Mediated Idea Sharing

Paper Title

GraftMind: Facilitating Group Ideation with AI-Mediated Idea Sharing

Publication Info

• Topic area: AI-mediated systems for enhancing group ideation
• Keywords: group brainstorming, individual ideation, AI mediator, synergy, evaluation apprehension, creativity stimulation, proactive assistance, idea sharing, semantic graph, dual-pathway theory

Background and Problem

• Problem / challenge: Traditional group brainstorming introduces evaluation apprehension, inhibiting creativity, while individual brainstorming lacks synergy. Existing hybrid methods alternate between the two modes but fail to integrate their strengths simultaneously.
• Significance: Combining the advantages of group and individual brainstorming could enhance creativity, reduce social inhibition, and improve ideation outcomes.
• Motivation and related work: Prior research has explored AI ideation partners and mediators but has not addressed real-time idea sharing during individual brainstorming. This paper builds on the dual-pathway theory of creativity stimulation to propose a new ideation setting.

Solution

• Proposed approach: GraftMind, a system enabling private ideation workspaces with AI-mediated real-time idea sharing based on collective ideas.
• Novelty:
  1. Introduces a novel group ideation setting that integrates the strengths of group and individual brainstorming.
  2. Develops an AI mediator offering three types of assistance: persistence, flexibility, and diversification recommendation.
  3. Embeds assistance into a digital whiteboard interface using a carrier pigeon metaphor for low-interruption delivery.
  4. Validates the system through a user study comparing ideation performance, synergy, and evaluation apprehension.
• Procedure and key techniques:
  • Semantic graph-based inference of users' ideation states using Qwen3-Rerank for text similarity.
  • Generation of assistance types (persistence, flexibility, diversification) using GPT-4.1 for abstraction and conceptualization.
  • Proactive delivery of assistance triggered by pauses in user activity, integrated into the digital whiteboard interface.

Results

• Concrete findings:
  • Experimental groups using GraftMind generated significantly more ideas (30–40 ideas per group) and idea clusters (15–25 clusters) than control groups (20–30 ideas, 10–20 clusters).
  • Evaluation apprehension scores were significantly lower in the experimental condition (U = 141.5, p < 0.001).
  • Synergy perception ratings were comparable between experimental and control groups (t(58) = −1.02, p = 0.156).
• Advantage over baselines:
  • Higher idea quantity and quality compared to conventional group brainstorming.
  • Reduced evaluation apprehension while maintaining synergy.
• Experiments / evaluation:
  • Between-subjects study with 60 participants (28 undergraduates, 32 graduates) divided into 10 groups per condition.
  • Metrics: idea quantity, originality (semantic clustering), ideation activity, synergy perception, evaluation apprehension.
  • Interaction logs and satisfaction questionnaires provided additional insights.
• Limitations and future work:
  • Non-randomized control condition and limited sample size.
  • Pause-based triggers for assistance timing may not fully capture cognitive states.
  • Need for personalized assistance mechanisms and large-scale comparative experiments.

Summary

GraftMind introduces a novel group ideation setting that combines the strengths of group and individual brainstorming by enabling private workspaces with AI-mediated real-time idea sharing. The system uses semantic graphs and generative models to infer users' ideation states and deliver three types of assistance (persistence, flexibility, diversification) through a low-interruption interface. A user study demonstrated its effectiveness in increasing idea quantity and quality, reducing evaluation apprehension, and maintaining synergy. Future work will address limitations in study design, timing mechanisms, and personalization to further refine this approach.