Introduction
Generative AI has democratized the creation of high-quality visual content, yet the standard interaction paradigm—a linear chat interface—often traps users in a "slot machine" workflow. Users provide a prompt, receive a result, and frequently settle for the first "good enough" output. This phenomenon, known as design fixation, prevents creators from exploring the full breadth of a creative space.
In this paper, researchers introduce HAIExplore, a system designed to break this cycle by explicitly scaffolding the two fundamental phases of creativity: divergent thinking (brainstorming) and convergent thinking (refinement). By moving away from the "chat-and-hope" model, HAIExplore helps users overcome the "Gulf of Envisioning" and produces more novel, high-quality results.
The Problem: Premature Convergence and the Gulf of Envisioning
Current human-AI collaboration suffers from two primary hurdles:
- Premature Convergence: The ease of generating a final image encourages users to stop exploring conceptual directions too early.
- The Gulf of Envisioning: Even when a user has a specific refinement in mind, translating that mental vision into the technical, often opaque language required by prompt-based AI models is difficult. This creates a barrier between intent and execution.
System Design: The HAIExplore Workflow
Grounded in Wallas’s model of creativity, HAIExplore splits the co-creation process into two distinct, structured stages:
Stage 1: Divergent Thinking (Brainstorming)
Instead of jumping straight to image generation, HAIExplore encourages users to explore conceptual "Idea Cards." The system utilizes associative thinking—leveraging LLMs to draw connections from remote domains like mythology, pop culture, or memes—to spark diverse directions. Users can interact with these cards, expanding on interesting concepts or "sparking" them to generate initial visual drafts. This stage ensures a wide net is cast before any specific design is finalized.
Stage 2: Convergent Thinking (Refinement)
Once a concept is selected, the system shifts to a refinement mode that prioritizes recognition over recall. Rather than forcing users to guess which keywords will modify an image, HAIExplore acts as a translator. It externalizes the AI’s internal parameters into a "Sketch"—a set of interpretable controls and dropdown menus. For example, if a user wants to change a character's mood, the system provides specific, selectable options rather than requiring a new, complex prompt. This reduces the cognitive load and allows for precise, iterative refinement.
Evaluation and Results
The researchers evaluated HAIExplore through a within-subjects user study with 12 participants, comparing it against a standard ChatGPT interface across three poster design tasks. The results demonstrated significant improvements across several key dimensions:
- Creativity Support: Using the Creativity Support Index (CSI), HAIExplore scored significantly higher in Enjoyment (16.83 vs. 10.25) and Exploration (17.25 vs. 9.17). Participants felt the system actively encouraged them to try new things rather than settling.
- Artifact Quality: Independent expert evaluators rated the images produced with HAIExplore as significantly more novel (4.87 vs. 4.02 on a 7-point scale) compared to those generated via standard chat.
- Usability: The system’s usability, measured via UMUX-Lite, was rated substantially higher (84.03 vs. 65.97), suggesting that structured scaffolding is more user-friendly than open-ended prompting.
Key Insights: The Ownership Paradox and Learning
One of the most intriguing findings was the "Agency-Ownership" trade-off. While users preferred HAIExplore for its results and ease of use, some reported a higher sense of personal "ownership" over ChatGPT results. Because the standard chat interface is more difficult to master, users felt the final result was more "theirs" due to the struggle involved in achieving it.
However, HAIExplore provided a unique educational benefit: scaffolded learning. By exposing the "hidden logic" of how prompts are constructed through its parameter-based refinement stage, users actually learned how to be better prompt engineers. Participants reported significantly higher self-perceived learning (5.25 vs. 2.50) and began transferring the strategies they saw in HAIExplore back to their standard AI interactions.
Limitations and Future Directions
While HAIExplore represents a significant step forward, the authors note several limitations. The study was conducted with a small, relatively homogenous sample of university students. Furthermore, the system is currently optimized for static image generation. Future work will explore applying this "brainstorm-then-refine" paradigm to more complex domains like video production and UI design, as well as incorporating direct manipulation (pointing and clicking) to further bridge the Gulf of Envisioning.
Conclusion
HAIExplore demonstrates that by moving beyond the chat box and explicitly scaffolding the creative process, we can move from AI as a "slot machine" to AI as a true creative partner. By supporting both the expansive nature of brainstorming and the precision of refinement, such systems not only produce better art but also help users become more capable creators.
Summary updated