When a user interacts with a Large Language Model (LLM), the speed of the model is often its greatest liability. In traditional creative workflows, the transition from an abstract goal to a concrete artifact is slow, allowing for a period of "divergent thinking" where multiple paths are explored. However, as Hariharan Subramonyam and colleagues describe in the "Gulf of Envisioning," the immediate delivery of a high-fidelity first draft by an AI can cause users to anchor their expectations to that specific result. This phenomenon, known as design fixation, limits creative exploration and prevents users from iterating toward more diverse or higher-quality alternatives.
The Mechanism of Cognitive Anchoring
Design fixation occurs because human cognition is naturally biased toward efficiency. We tend to recognize the primary or typical functions of objects—a concept Tony McCaffrey calls the "Obscure Features Hypothesis"—and struggle to look beyond established mental models once a problem is framed.
In the context of AI, this bias is exacerbated by the "fidelity" of the output. When a tool provides a functional-looking solution immediately—whether it is a block of code in "vibe coding" workflows or a polished image—users often redirect their energy toward refining that specific version rather than evaluating its fundamental structure. Daeun Choi notes that highly polished visual outputs stifle creativity by appearing as "final" solutions, leaving little cognitive room for the user to project their own imagination or explore "what if" scenarios.
This anchoring effect isn't limited to novices. Alistair Shilton and Santu Rana identify cognitive entrenchment in experts, who often over-exploit known successful designs through incremental adjustments. Paradoxically, Susan P. Imberman suggests that deep domain expertise can create a "plausibility gap," where professionals limit their solutions to current technological constraints, while less-experienced students remain open to more radical, imaginative leaps.
Mapping the Design Space
To break this fixation, researchers are shifting the interaction paradigm from "prompt-and-response" to structured design space exploration. Instead of asking for a single answer, tools like Luminate, developed by Sangho Suh and Haijun Xia, task the LLM with first identifying the underlying dimensions of a task (such as tone, genre, or setting). By generating a diverse array of responses across these dimensions simultaneously, the system forces the user to see the "extremes" of the design space, encouraging lateral thinking.
A similar approach is taken in scientific ideation. Daniel S. Weld and Tom Hope introduced SCIDEATOR, which breaks research papers into structural "facets"—Purposes, Mechanisms, and Evaluations. By recombining these facets from distant research areas, the system helps researchers break out of "cognitive rifts" and discover cross-disciplinary analogies that they might otherwise overlook due to their own niche expertise.
Introducing "Alien" Friction
If human creativity is constrained by a "cultural horizon"—the tendency to combine concepts that are already familiar or culturally adjacent—then the role of AI should be to provide "alien" perspectives. James Evans and Jamshid Sourati demonstrate that AI can be tuned to generate hypotheses that are scientifically sound but socially "distant" from human expertise, effectively filling collective blind spots.
However, simply having a powerful model is not enough. Runlin Duan and colleagues found that LLMs often exhibit the same cognitive fixations as humans, relying on "safe" conceptual categories unless specifically prompted for diversity. Furthermore, providing human-generated examples through few-shot prompting can inadvertently anchor the model, causing it to mimic the specific conceptual boundaries of the samples rather than expanding the design space.
Design Patterns for Divergence
To mitigate these cognitive traps, several design interventions have emerged:
- Simultaneous Diversity: Presenting multiple diverse outputs at once, rather than iteratively, to break the serial interaction loop and prevent premature convergence Subramonyam et al..
- Low-Fidelity Scaffolding: Using abstract keywords or low-fidelity sketches to maintain "interpretive flexibility," allowing the user to remain in a conceptual state longer Choi.
- Intentional Divergence: Building "Designer Models" that identify a user's repetitive habits and intentionally suggest options that challenge their established patterns Antonios Liapis, Georgios N. Yannakakis, and Julian Togelius.
- Prompt Recommendation: Shifting the user's burden from "recalling" what to ask to "recognizing" potential directions through suggested follow-up prompts, which increases expressiveness without raising cognitive load Jason Kim.
By intentionally introducing friction and structure into the generative process, these systems move AI from a tool of mere imitation to a true partner in exploration, helping creators navigate the tension between what is familiar and what is possible.