This paper investigates how the "sociotechnical configuration" of a digital workspace—specifically the difference between a 2D planar canvas and a 3D immersive VR environment—shapes how groups brainstorm alongside generative AI. The researchers found that while VR increases spatial presence and encourages more balanced participation, the specific format of AI output (long-form text vs. short keywords) is a primary driver of how groups evaluate and adopt AI-generated ideas.
TL;DR
A comparative study of 2D (Miro) and VR (Noda) workspaces reveals a "participation paradox": while immersive environments foster spatial presence and balanced group roles, they can also cause teams to "drift apart" and fragment. Furthermore, the verbosity of AI suggestions—dictated by the platform's display constraints—significantly influences whether a group perceives the AI as "insightful" or merely a source of generic keywords.
What Problem Does This Paper Address?
Remote collaborative ideation often lacks the "spark" of in-person sessions due to reduced spatial presence and limited non-verbal cues. While generative AI is increasingly integrated into these tools to boost creativity, we don't yet understand how the environment (2D vs. 3D) and the AI's presentation (long sentences vs. short keywords) interact to change group dynamics.
The authors argue that these platforms are not just neutral tools but "sociotechnical configurations" where spatial mediation, interaction modality, and AI output format co-vary. They aim to understand how these bundles of features affect attention distribution, the negotiation of authority, and the collective uptake of AI suggestions.
Methods
The researchers conducted a between-subjects study with 14 all-women triads (42 participants). They compared two commercially available platforms:
- Miro (Planar): A 2D digital whiteboard where AI (GPT-4o-mini) generates detailed, full-sentence suggestions.
- Noda (Immersive VR): A 3D VR workspace where AI (GPT-3.5 Turbo) generates concise, keyword-style nodes.
Teams followed a structured "brainwriting" task:
- Subtask 1: Human-only idea generation.
- Subtask 2: AI-supported expansion (invoking the "brainstorm" features).
- Subtask 3: Group convergence (selecting three finalist ideas to "fund").
The team used a mix of video analysis to track collaboration patterns, post-task questionnaires (NASA-TLX for workload, MEC-SPQ for presence), and computational text analysis (Learning via Surprisability) to identify semantic differences in the ideas generated.
Datasets
The study generated a new idea corpus:
- Total ideas: 1,260 distinct ideas.
- Sources: 308 human-authored, 937 AI-generated, and 15 group-authored (ideas edited or combined by the team).
- Evaluation: Participants rated three mutually exclusive sets of ideas on relevance, innovation, and insightfulness.
Key Insights and Ideas
- The Participation Paradox: In the VR environment, the "spatial freedom" that allowed participants to explore also led to them "drifting apart." Teams often worked in separate corners, facing away from each other, which fragmented group coherence.
- Verbosity as a Design Variable: AI output format was a major differentiator. Miro’s AI averaged 21 words per idea, while Noda’s averaged only 3 words. Participants rated the longer Miro suggestions as significantly more "insightful" and "innovative," and selected them as finalists far more often (10/21 in Miro vs. 3/20 in Noda).
- The "Invisible" AI Partner: In both platforms, AI was invoked individually and asynchronously. There were no mechanisms to make an AI invocation a "group event," meaning participants often had "uneven awareness" of what the AI was contributing to the shared space.
- Balanced Roles in VR: Despite the fragmentation, VR was associated with more balanced role distribution during the final selection phase. Five out of seven Noda groups shared "navigator" and "driver" roles equally, compared to only three out of seven in Miro.
Key Results
- Idea Quantity: AI significantly multiplied the idea pool, tripling it in Miro and quintupling it in Noda.
- Idea Quality: In both settings, ideas that groups authored with AI available were rated higher than the ideas they proposed alone.
- Presence & Workload: As expected, VR (Noda) provided a much stronger sense of "being there" but came with significantly higher physical demand (Mdn = 1.9 vs 1.0 on a 7-point scale).
- Semantic Differences: AI-generated ideas tended to use more "programmatic" language (e.g., promote, practice, organ, host), while human ideas were more situated and concrete (e.g., hall, sus).
Limitations and Future Work
The authors acknowledge several constraints:
- Sample Homogeneity: The study used all-women undergraduate groups to reduce gender-related communication variability, but results may not generalize to mixed-gender or professional teams.
- Confounded Variables: Because they used "off-the-shelf" platforms, the AI models (GPT-4o-mini vs. GPT-3.5) and text-entry speeds (physical keyboard vs. VR pointer) differed. It is hard to say if Miro's success was due to the better model or just the ease of reading long text.
- Short-term Focus: The study doesn't capture how teams might adapt to these tools over weeks or months.
Future Work: The authors suggest designing "phase-sensitive" workspaces—providing spatial freedom during brainstorming but using "avatar pull-toward gestures" or shared focal points to draw teams together during decision-making. They also call for "group-visible" AI invocation to ensure the AI acts as a true team member rather than a private assistant.
Summary updated