Design Fixation: Cognitive Constraints in Design

• Design fixation is a cognitive phenomenon where designers become anchored to existing examples, limiting the exploration of novel solutions.
• Empirical studies use metrics like the Design Fixation Score (DFS) and novelty ratios to quantify repetitive features and constrained output.
• Intervention strategies such as design-by-analogy and generative friction actively disrupt fixation, promoting diverse and creative ideation.

Design fixation is a cognitive and behavioral phenomenon in which designers become anchored to existing examples, initial concepts, or surface features, thereby constraining the exploration of novel or diverse solutions. This constraint manifests as premature convergence, limited idea variety, and recursive elaboration of familiar patterns, often exacerbated in human–AI co-creation environments and under template-driven workflows. Mounting empirical and theoretical analyses across design research, HCI, and creativity support tool (CST) literature situate fixation both as an individual cognitive bias and as an emergent property of collaborative, technologically mediated practices.

1. Conceptual Foundations and Cognitive Mechanisms

Design fixation is consistently defined as “(blind) adherence to a set of ideas or concepts limiting the output of conceptual design,” a formulation that traces to Jansson and Smith (1991) and is operationalized in both lab studies and field research (Li et al., 10 Feb 2026, Wen et al., 20 Dec 2025, Wadinambiarachchi et al., 2024, Parsons et al., 2021). At the cognitive level, fixation arises due to over-reliance on familiar mental schemas and surface analogies, activating repetitive retrieval pathways in memory (Simon, 1973; Guilford, 1967; Holyoak & Thagard, 1996). Eye-tracking and neuroimaging confirm that superficial similarity cues crowd out deeper, relational or structural mappings, inhibiting divergence and abstract transfer (Li et al., 10 Feb 2026). In generative AI contexts, this phenomenon generalizes to model-internal patterns—i.e., statistical convergence on high-frequency patterns—so that both humans and GenAI systems exhibit constrained output distributions (Chen et al., 9 Feb 2025). Formally, if $\mathcal G$ denotes a generative model mapping prompts $p \in \mathcal P$ to designs $d \in \mathcal D$, fixation manifests as $\Pr(\mathcal G(p) \in \Delta) \gg 1/|\mathcal D|$ for a narrow region $\Delta$.

2. Manifestations and Empirical Metrics

Fixation is empirically observed across modalities—text, image, workflow design, and beyond—and quantified via diverse operational metrics. In image generation, the Design Fixation Score (DFS) is computed as the proportion of salient features in an output that overlap with an initial exemplar (Wadinambiarachchi et al., 2024). The variety and originality of generated ideas are measured by metrics such as the average pairwise cosine distance of CLIP or other embedding-space representations and by the count of unique conceptual clusters (“image clusters”) (Wen et al., 20 Dec 2025, Chen et al., 9 Feb 2025). For example, presence of fixating features in sketches after Midjourney or Image Search exposure resulted in DFS ≈ 0.35 versus DFS ≈ 0.25 (baseline), with strong Bayesian evidence for increased fixation (Bayes factors > 100) (Wadinambiarachchi et al., 2024). Fixation also manifests in iterative workflows as high local refinement ratios, adoption of default options, and low cluster count (premature convergence) (Wen et al., 20 Dec 2025).

In text generation, recurring high-frequency terms, surface-level descriptions, and invariant context indicate fixation (Chen et al., 9 Feb 2025). For instance, novelty ratios—defined as $P_\mathrm{novelty} = U/(U + S)$—were 67.4% for ChatGPT compared to 78.2% for a curated set of Red Dot design entries. In group design, fixation is quantified by the Fixation Score (FS), the proportion of ideas duplicating the initial example; groupthink and lock-in around AI-generated starting points are recurrent themes (Lee et al., 2024).

3. Causal Factors and Exacerbating Conditions

Design fixation is triggered and reinforced by several factors across human and AI-mediated workflows:

• Input–Output Pipelines in AI Tools: Simplified prompt→output flows in foundation models (e.g., LLMs, image generators) privilege efficiency over exploration. Training on large but biased datasets, these systems amplify statistically dominant patterns, reducing both human and algorithmic opportunity for divergent ideation (Li et al., 10 Feb 2026, Chen et al., 9 Feb 2025).
• Template and Example Use: Example lists and software chart templates induce priming and local hill-climbing, anchoring user trajectories in the problem space and promoting feature copying (Chan et al., 2024, Parsons et al., 2021).
• Surface Feature Salience: Reliance on surface similarity, such as color, shape, or style, overrides relational or functional analogies, perpetuating shallow recombination rather than transfer or abstraction (Jiang et al., 2021, Li et al., 10 Feb 2026).
• Human Factors: Emotional attachment, sunk cost (effort investment), client influence, and habitual reuse further entrench fixation, as do over-familiarity with technical labels or strict adherence to “best practices” (Parsons et al., 2021, Wadinambiarachchi et al., 2024).
• AI-Specific Biases ("GenAI Design Fixation"): Saillient in design-focused GenAI, output diversity is constrained by model bias, training set imbalances, and limited prompt expressivity, resulting in attribute homogeneity, repetitive motifs, and resistance to novel elements even under diverse prompts (Chen et al., 9 Feb 2025).

4. Intervention Strategies and Tool Design

Mechanisms to counter design fixation target both representational and process-level factors:

• Design-by-Analogy (DbA): Embedding analogy-driven processes throughout the creative workflow mitigates fixation by exposing designers to relationally structured, far-domain stimuli. Six forms of representation—semantics/text, visual/appearance, material/structure, function/attribute, interaction/workflow, unconventional contexts—can each be leveraged to break surface-level anchoring by broadening the mappable knowledge space (Li et al., 10 Feb 2026, Jiang et al., 2021).
• Process-Oriented Co-Creation: Scaffolding both divergent (brainstorming) and convergent (refinement) stages, with explicit interfaces for ideation branching, associative prompting, and parametric refinement, promotes exploration and reduces premature convergence (Wen et al., 20 Dec 2025). For example, the HAIExplore system achieved higher novelty scores ($4.87 \pm 0.62$ vs. $4.02 \pm 0.54$, $p = 0.004$) and trends toward increased ideation branching relative to a standard chat-based AI interface.
• Generative Friction: Introducing intentional disruptions to “seamless” AI output—fragmentation of text, deliberate delivery delays, semantic ambiguity—recasts outputs as semi-finished, inviting appropriation and reinterpretation (Kocaballi et al., 29 Mar 2026). The efficacy of such friction is moderated by “Friction Disposition,” denoting a user’s tolerance for ambiguity and propensity to treat resistance as invitation.
• Contextualization and Model-building: Contextualized example presentation (e.g., in-space overlays) encourages users to formulate internal models of the problem space, supporting broader exploration. List presentations, by contrast, induce local hill-climbing and increased fixation (Chan et al., 2024). Odds ratios for hill-climbing behavior under list vs. context conditions exceed six.
• Devil’s Advocate Conversational Agents: LLM-based agents employing dynamic, retrieval-augmented counterarguments, privilege critical reflection, disrupting groupthink-driven fixation. Monitoring conversational diversity and triggering timely dissent increases the proportion of unique ideas and raises critical reflection indices (Lee et al., 2024).

5. Formal and Quantitative Characterizations

A range of formal models have been leveraged to analyze and operationalize design fixation:

Metric / Model	Definition / Role	Source Paper(s)
DFS (Design Fixation)	Proportion of repeated features from exemplar	(Wadinambiarachchi et al., 2024)
FS (Fixation Score)	$$1 - (|\text{UniqueNewIdeas}| / |\text{TotalIdeas}|)$$	(Lee et al., 2024)
Novelty Ratio	$p \in \mathcal P$0 (unique/shared keywords)	(Chen et al., 9 Feb 2025)
Diversity (D, $p \in \mathcal P$1)	Average pairwise embedding distance	(Wen et al., 20 Dec 2025, Chen et al., 9 Feb 2025)
Gentner’s Mapping	$p \in \mathcal P$2	(Li et al., 10 Feb 2026)
Hill-Climbing Likelihood	$p \in \mathcal P$3	(Chan et al., 2024)

CSTs are further evaluated on fluency, originality, and typicality metrics, as well as via user-reported awareness scales and post-task qualitative coding.

6. Domain Applications and Risks

Fixation and its mitigation feature across multiple domains:

• Creative Industries: Tools such as Umitation (UI behavior retargeting) and VST (flexible style transfer) decompose domain-specific patterns and support analogical leaps beyond dominant aesthetics (Li et al., 10 Feb 2026).
• Manufacturing: BioSpark, InnoGPS, and cross-domain fabrication repositories surface alternative process paradigms, reducing parametric tweaking and breaking from legacy fixation (Li et al., 10 Feb 2026).
• Education and Services: Systems leveraging multimodal ontologies and AR analogies scaffold meta-cognitive reflection, challenging terminological grooves and promoting continual learning (Li et al., 10 Feb 2026).
• Visualization Practice: Structured experimentation, cross-domain inspiration, and critique loops are empirically validated to disrupt fixation in data visualization workflows (Parsons et al., 2021).

Key risks include (i) the erosion of innate creative skills through over-reliance on analogical AI; (ii) oversimplification of real-world complexity by standardized representations; and (iii) propagation of data biases through analogy mapping or generative pipelines (Li et al., 10 Feb 2026).

7. Future Directions and Open Issues

Research continues to examine the persistence and mitigation of fixation in both human and AI systems:

• Architectural and Data Remedies for GenAI: Optimizing model objectives for output diversity, utilizing balanced training corpora, and dynamically steering generation via diversity metrics are identified as high-priority directions (Chen et al., 9 Feb 2025).
• Evaluation and Benchmarking: Convergent validation of CST efficacy requires expansion of benchmark datasets, standardization of fixation/deliberation metrics, and rigorous user study designs crossing modalities (Jiang et al., 2021).
• Personalization and Adaptation: Adapting system interventions (e.g., friction, timing, representation) to user-specific profiles, including ambiguity tolerance and workflow preferences, is recognized as critical for widespread efficacy (Kocaballi et al., 29 Mar 2026, Li et al., 10 Feb 2026).
• Sociotechnical and Ethical Considerations: Addressing the downstream effects of analogical bias, equitable distribution of creative agency, and avoidance of “ossified” solution spaces through AI mediation (Li et al., 10 Feb 2026).

In synthesizing evidence across recent arXiv and HCI scholarship, the design fixation construct emerges as a central concern in both the theory and praxis of creativity support. Its operationalization, measurement, causal understanding, and mitigation are tightly coupled not only with cognitive and behavioral science but also with the representational, architectural, and interactional properties of contemporary AI and CST systems.