Ask someone to design a logo using an AI image generator and watch what happens. They type a prompt, receive an image, adjust the prompt slightly, receive a similar image, and iterate within an increasingly narrow visual neighborhood. Within minutes, they have converged on a solution that is competent but unremarkable — and they have stopped exploring. This pattern, called design fixation, is not a failure of the user or the AI but a failure of the interaction paradigm itself. The standard prompt-and-receive cycle actively suppresses the divergent thinking that creative work requires.
The Fixation Evidence
Wadinambiarachchi, Kelly, Pareek et al. (2024), in a CHI paper that has accumulated 141 citations, provide the most direct experimental evidence. In a between-participants study with 60 participants performing a visual ideation task, they found that using an AI image generator during ideation leads to higher fixation on an initial example — not lower. Participants who worked with AI produced fewer ideas, with less variety and lower originality compared to a baseline group working without AI assistance.
The qualitative analysis reveals why. The AI's polished outputs anchor the user's thinking. When a generator produces a high-quality image from a rough prompt, the user's conceptual exploration shifts from "what could this be?" to "how do I refine this?" — a transition from divergent to convergent thinking that happens too early in the creative process. The richness and completeness of AI outputs, paradoxically, narrows the creative space by providing a local optimum that feels sufficient.
The effectiveness of co-ideation depends critically on how participants approach prompt creation and how they respond to AI suggestions. Some users treat AI outputs as starting points for exploration — deliberately seeking surprising or unexpected results. Others treat them as drafts to be polished — converging immediately on the most promising output. The former approach preserves creative range; the latter collapses it.
Structuring the Creative Process
Wen, Phung, Mehrotra et al. (2025), from the Max Planck Institute for Software Systems and Microsoft, propose a solution grounded in Wallas's classical model of creativity. Their insight is that creative work has distinct phases — preparation, incubation, illumination, verification — and that current AI interaction paradigms collapse these phases into a single prompt-response loop.
Their system, HAIExplore, restructures the human-AI interaction into explicitly separated stages. A divergent thinking stage scaffolds high-level conceptual exploration: users brainstorm ideas in a conceptual space before generating any images, mapping the landscape of possibilities rather than committing to a specific direction. A convergent refinement stage then scaffolds low-level exploration of variations: users work with interpretable parameters and options that externalize their refinement intentions, making the convergence process controllable and transparent.
In a within-subjects study comparing HAIExplore with ChatGPT's linear chat interface for creative image generation, the structured paradigm measurably mitigates design fixation. Users explore more of the conceptual space, consider more diverse approaches, and report greater perceived control over the creative process. The key mechanism is temporal separation: by preventing users from seeing polished outputs during the brainstorming phase, the system keeps the divergent thinking space open longer.
Ambiguity as Creative Fuel
Dalsgaard (2025) contributes a complementary theoretical perspective: the role of creative ambiguity and cognitive tension in generative AI tools. The argument is that productive creativity arises not from clarity but from the tension between competing interpretations — and that well-designed AI tools should preserve and amplify this tension rather than resolving it prematurely.
Current AI systems are optimized for resolution. They produce clear, coherent, complete outputs. But the creative process thrives on productive ambiguity — the moment when an idea is half-formed and could develop in multiple directions. A sketch that is deliberately imprecise invites interpretation. A sentence fragment suggests possibilities that a polished paragraph forecloses. If AI tools resolve ambiguity too quickly — by producing finished outputs from vague prompts — they shortcut the cognitive tension that drives creative discovery.
The design implication is counterintuitive: AI creative tools should sometimes produce deliberately incomplete, ambiguous, or contradictory outputs. A brainstorming tool that generates conceptual tensions ("what if the logo were both organic and geometric?") rather than resolved images may better serve the divergent thinking phase than a system that generates polished results from the start.
Redesigning the Loop
These three perspectives converge on a structural critique: the standard generative AI interaction paradigm — type prompt, receive output, refine prompt — is poorly matched to how creativity actually works. Creative thinking requires divergent exploration before convergent refinement, tolerance for ambiguity before resolution, and conceptual breadth before visual specificity. Current tools provide the opposite: immediate resolution, rapid convergence, and premature specificity.
The solution is not better AI models but better interaction architectures. The models already produce outputs of sufficient quality. The problem is that these high-quality outputs arrive too early in the process, anchoring exploration before it has begun. Systems that scaffold the creative process — by separating divergent and convergent phases, by delaying resolution, by generating conceptual provocations rather than finished products — can turn AI from a fixation engine into an exploration engine. The creative potential of generative AI will be realized not by improving what the AI produces but by restructuring when and how it produces it.