Stanford CS547 HCI Seminar | Autumn 2025 | Tracing and Shaping Paths in Design Space

Key Concepts

• AI-supported creativity: Investigating how AI tools (CSTs) impact the human creative process, moving beyond simple evaluation of outputs to analyze the process itself.
• Design Space & Trajectories: Mapping user navigation within the vast possibility space of creative artifacts, analyzing “trajectories” of interaction.
• Homogenization of Ideas: The tendency of AI tools, particularly large language models like ChatGPT, to generate more similar ideas among users compared to analog methods.
• Multi-Actor Fuzzy Linkography: A novel visualization technique for analyzing influence dynamics between humans and AI during ideation, revealing patterns of interaction and potential limitations.
• The Importance of Interaction Design: Recognizing that the interface and interaction patterns significantly shape the creative outcome when using AI tools.

Bridging AI and HCI: Understanding Creative Processes

The research presented focuses on understanding and improving computer-supported creative processes, bridging the fields of Artificial Intelligence and Human-Computer Interaction. The core goal is to develop empirical methods for evaluating creativity processes – divergence, convergence, inspiration, blockage – rather than simply quantifying the resulting artifacts. This involves analyzing user “trajectories” through “design space,” the totality of possible configurations within a creative domain.

Early CST Research & Methodological Development

Initial work involved developing and analyzing various Creativity Support Tools (CSTs). Germinate, a mixed-initiative casual creator for rhetorical game design, used Answer Set Programming to generate playable game prototypes. Meet the Ganimals explored user navigation of a latent space through GAN-based image evolution. Redactionist, an eraser poetry tool, facilitated “Expressive Range Coverage Analysis” (ERA), characterizing the range of artifacts a generative model can produce. Patchwork, an AI-supported worldbuilding canvas within Midjourney, presented challenges due to its large and complex design space, analyzing 6,424 image prompt sequences. A comparative study of ChatGPT versus Brian Eno’s Oblique Strategies cards highlighted the potential for AI to lead to more homogeneous ideas despite potentially higher scores on metrics like flexibility, fluency, and elaboration. Key analytical frameworks developed include Linkography (originally architectural) and its automated adaptation, Fuzzy Linkography, used to identify “design moves” and reveal user behavior patterns. Analysis of linkographic data also revealed identifiable Designer Archetypes – patterns of user behavior within CSTs.

The Homogenization Effect of ChatGPT

A central finding of the research is that ChatGPT, while performing well on traditional creativity metrics, leads to more homogeneous idea generation among users compared to the more analog approach of using Oblique Strategies cards. This effect was investigated using a novel methodology: Multi-Actor Fuzzy Linkography. This technique parses transcripts of human-AI interactions into structured entries, creates links between ideas based on semantic similarity, and colors these links to denote the actor initiating the idea (red = human, blue = AI, purple = interaction).

Analyzing Interaction Dynamics with Linkography

The linkography revealed several key patterns. AI frequently generated batches of irrelevant ideas (“misfires”) that were ignored by users, and users often became stuck in repetitive reprompting cycles. The analysis demonstrated a predominantly one-way influence from the AI to the user, with the AI rarely building upon human ideas. Users increasingly shifted to a “curationist mode,” focusing on filtering and integrating ChatGPT’s output rather than generating original ideas, particularly over time. Quantitative analysis of backlink densities confirmed these observations, showing stronger semantic relatedness between ChatGPT-generated ideas than between human-generated ideas.

Expanding Analytical Applications

Beyond the ChatGPT/Oblique Strategies study, the linkography technique was applied to other datasets. The Autotographer project utilized CLIP embeddings to spatialise museum data, allowing users to navigate and collect images, providing a new platform for analyzing creative trajectories and revealing a preference for exploring the edges of semantic space. A preliminary application of linkography was also used to visualize the career paths of researchers using Google Scholar title analysis.

Implications and Future Directions

The research suggests that while AI tools like ChatGPT are powerful, their current design can stifle creativity by promoting homogenization and user passivity. The interaction design is crucial, and there is value in the “lost poetry” – the nonsensical or “wrong” outputs that can push creative boundaries. The development of multi-actor fuzzy linkography provides a valuable analytical tool for understanding these complex interaction dynamics. Future work aims to leverage this data to build more proactive and responsive CSTs, balancing intervention with user agency and fostering more diverse and innovative creative outcomes.