PaperlayerView exact revision on arXiv

Source-linked AI summary

The Effects of Generative AI on Design Fixation and Divergent Thinking

Samangi Wadinambiarachchi, Ryan M. Kelly, Saumya Pareek, Qiushi Zhou, Eduardo Velloso

arXiv:2403.11164v1cs.HC

TL;DR

Generative AI has been proposed to augment creativity, but empirical evidence about its effects during design ideation remains limited. This study evaluates AI-generated images as inspiration in a visual ideation task and finds higher fixation alongside poorer divergent-thinking outcomes than baseline support.

  • Problem

    Empirical evidence is limited on how generative AI affects design fixation and divergent thinking during ideation, including how different interactions with AI image generators influence effectiveness.

  • Method

    The study evaluated AI-generated images as inspiration in a visual ideation task using design fixation, fluency, variety, and originality measures, with Bayesian multilevel models for analysis.

  • Results

    AI support produced higher design fixation and lower fluency, variety, and originality than baseline conditions, with fixation also shifting from the initial example to AI-generated images.

  • Takeaways & Limitations

    For novice designers in a fixed-time visual ideation task, sketching may be more productive than seeking inspiration through AI, while prompt creation and responses to AI suggestions remain redesign opportunities.

  • Takeaways & Limitations

    The findings primarily reflect short-term, somewhat naïve AI use in a rapid 20-minute task and do not establish how effects vary with professional design expertise.

Abstract

from arXiv · show

Generative AI systems have been heralded as tools for augmenting human creativity and inspiring divergent thinking, though with little empirical evidence for these claims. This paper explores the effects of exposure to AI-generated images on measures of design fixation and divergent thinking in a visual ideation task. Through a between-participants experiment (N=60), we found that support from an AI image generator during ideation leads to higher fixation on an initial example. Participants who used AI produced fewer ideas, with less variety and lower originality compared to a baseline. Our qualitative analysis suggests that the effectiveness of co-ideation with AI rests on participants' chosen approach to prompt creation and on the strategies used by participants to generate ideas in response to the AI's suggestions. We discuss opportunities for designing generative AI systems for ideation support and incorporating these AI tools into ideation workflows.

1 INTRODUCTION

The paper examines whether AI-generated images support creativity during visual ideation, addressing limited empirical evidence about their effects on fixation and divergent thinking. A between-participants experiment compares no inspiration, image search, and GenAI support.

  • Design fixation occurs when exposure to one idea anchors subsequent ideas and restricts exploration of the design space.The paper describes fixation as blind adherence to ideas or concepts that limits conceptual design output.
  • Generative AI image tools were promoted as potentially augmenting creativity by rapidly producing visual inspiration from user prompts.The paper notes that the empirical effects of these tools during design tasks remained unclear.
  • The experiment compared no inspiration, Google Image Search, and Midjourney while participants sketched chatbot-avatar ideas.The study used a between-participants visual ideation task.
  • The research questions asked how AI-generated images affect fixation and divergent thinking, and how interaction styles influence ideation effectiveness.These questions frame both the comparative experiment and the qualitative analysis of AI interaction.
  • AI-supported participants showed higher fixation and lower fluency, variety, and originality than the baseline condition.The qualitative analysis linked fixation to prompt creation and idea generation in response to AI images, including fixation displacement onto AI outputs.
  • The study contributes empirical evidence on how AI-generated images influence design fixation and divergent thinking measures.It also identifies opportunities to redesign AI ideation support and mitigate fixation.

2 RELATED WORK

Prior research treats design fixation as a major constraint on creative exploration and has examined ways to reduce it. Generative AI was proposed as inspiration support, but HCI evidence about its effects remained nascent.

  • Design Fixation: Design fixation restricts ideation by encouraging adherence to preknown ideas, producing narrower and potentially less original design outcomes.It has been studied across cognitive science, design, education, engineering, and psychology.
  • Design Fixation: Exposure to example solutions is associated with fixation, although effects on creativity vary with inspiration modality, abstraction, and designer expertise.Classic experiments compared participants given example designs with controls receiving no examples.
  • Design Fixation: The literature distinguishes unconscious adherence, conscious blocking, and intentional resistance as different fixation effects.These range from copying features without awareness to deliberately favoring familiar solution areas.
  • Overcoming Fixation: Proposed anti-fixation strategies include prototyping, reminders, lateral-thinking methods, incubation periods, intelligent agents, and design by analogy.The reviewed approaches target different stages of idea generation and exploration.
  • Research Gap: This study adapts established design-fixation methods to add empirical HCI evidence about AI image generators during visual design tasks.The paper specifically measures copying-based fixation alongside divergent-thinking outcomes.
  • Generative AI and Creativity: Generative AI can produce diverse visual stimuli through natural-language prompts and has been proposed as a creativity-support or co-creative tool.Prior work includes AI assistants intended to help ideation and overcome fixation, but findings remained early and exploratory.

3 METHOD

The study used a three-condition between-participants experiment in which participants sketched chatbot avatars after seeing a common example. Outcomes included fixation, fluency, variety, and originality, analyzed with Bayesian multilevel models.

  • Study Design: The independent variable was inspiration stimulus: Baseline, Google Image Search, or GenAI access during the task.The GenAI condition used Midjourney V4, accessed through textual prompts that generated four images per prompt.
  • Study Design and Materials: The example avatar contained 14 salient features used to quantify fixation through feature overlap in participant sketches.Participants saw the example without callouts, while two blind raters counted repeated features.
  • Study Design and Materials: Participants were asked to produce as many chatbot-avatar sketches as possible within 20 minutes.All conditions received the same brief, which included an example avatar for reference.
  • Measures: Fluency was the number of sketches produced, while variety captured solution-space coverage across sketch clusters.Variety was normalized so sketches in one cluster scored 0 and sketches spanning every cluster scored 1.
  • Measures: Originality measured how unusual each sketch was relative to other participants’ ideas, using cluster membership and its complement to 1.The measure was based on the number of other participants with ideas in the same cluster.
  • Participants and Procedure: The study recruited 60 participants, balanced conditions by gender, and conducted individual laboratory sessions lasting 45–60 minutes.Sessions included questionnaires, the main task, a post-study questionnaire, and a semi-structured interview.

4 RESULTS

The paper theorizes that inspiration stimulus affects fixation, fluency, variety, and originality, partly through time allocation during ideation. It further proposes links from fluency to variety and from variety to originality.

  • Theorized Causal Model: The theorized DAG treats inspiration stimulus as affecting design fixation, fluency, variety, and originality.These are the study’s principal creative-output outcomes.
  • Theorized Causal Model: The choice of inspiration stimulus is theorized to change time spent sketching rather than seeking inspiration, thereby affecting sketch fluency.Fluency is defined in the model as the number of sketches produced.
  • Theorized Causal Model: Higher fluency is proposed to increase variety because producing more sketches raises the likelihood of covering more of the solution space.The passage presents this as a likely relationship within the theorized model.
  • Theorized Causal Model: Greater variety is proposed to lead to more original ideas.This is stated as a likely relationship in the paper’s causal graph.

4.2 Design Fixation

The analysis models design fixation as overlap between participants’ sketches and salient features in the initial example. GenAI exposure was associated with higher fixation than baseline, with strong posterior support.

  • 4.2 Design Fixation: Design Fixation was modeled as the number of salient example features also appearing in each participant’s sketches.The model used a binomial distribution with N = 14 and included participant and image random effects.
  • 4.2 Design Fixation: 100% posterior probability indicated that GenAI increased Design Fixation, with mean = .32 and 89% CI [.11, .55].The Bayes Factor was 124, indicating extreme support for higher fixation than the No support baseline.
  • 4.2 Design Fixation: Higher Design Fixation scores indicate that a greater percentage of salient features from the example appeared in participants’ sketches.The figure defines higher scores as worse fixation.

4.3 Fluency

Fluency was analyzed as the number of sketches generated, using models that separated direct stimulus effects from effects mediated by time on task. Neither inspiration condition enhanced fluency over baseline.

  • 4.3 Fluency: Fluency was modeled as the number of sketches produced, using negative binomial models with and without time on task as a covariate.This accounted for different sketching times across the GenAI and Image Search groups.
  • 4.3 Fluency: Both Image Search and GenAI had detrimental total effects on Fluency compared with baseline.The GenAI total-effect estimate was -.21.
  • 4.3 Fluency: Neither Image Search nor GenAI enhanced Fluency compared with baseline, with both generally producing lower fluency.After controlling for total output time, Image Search had a minimal effect, while GenAI’s direct effect was relatively neutral.

4.4 Variety

Variety was modeled from the number of sketch clusters, separating total stimulus effects from direct effects after accounting for fluency. Neither Image Search nor GenAI improved variety over baseline.

  • 4.4 Variety: Variety was modeled as the number of clusters containing a participant’s sketches, minus one, using negative binomial models.The analyses estimated total effects and direct effects after including Fluency as a covariate.
  • 4.4 Variety: GenAI had a negative total effect on Variety, with mean = -.29 and 89% CI [-.75, .16].The model assigned an 86% probability to a negative effect, with Bayes Factor 5.93.
  • 4.4 Variety: Image Search had an even more negative total effect on Variety, with mean = -.40 and interval [-.87, -.07].Its Bayes Factor was 11.38, providing strong support for a negative effect.
  • 4.4 Variety: After accounting for sketch count, Image Search showed little direct effect, whereas GenAI retained a negative direct effect on Variety.The reported direct-effect estimates were .02 for Image Search and -.15 for GenAI.

4.5 Originality

Originality was modeled from how many other participants shared a sketch’s cluster, with direct and variety-mediated effects considered. Both stimuli showed small negative total effects on originality.

  • 4.5 Originality: Originality was modeled from the number of other participants with sketches in the same cluster, using direct and variety-mediated effects.The models included the inspiration stimulus and, for direct effects, Variety as a covariate.
  • 4.5 Originality: All parameter estimates converged with effective sample sizes well above 1000 and R-hat values of 1.00.The reported compatibility intervals are percentiles of posterior distributions, not frequentist confidence intervals.
  • 4.5 Originality: GenAI had a small negative total effect on Originality, with mean = -.03 and 89% CI [-.06, .00].The model assigned a 97% probability to a negative effect, with Bayes Factor 35.
  • 4.5 Originality: Image Search had a slightly less negative total effect on Originality, with mean = -0.01 and interval [-.04, .01].The Bayes Factor was 3.9, providing only anecdotal evidence against a positive effect.

4.6 Why did ideating with Generative AI cause design fixation?

AI-supported ideation appeared to induce fixation through prompts tied to the brief or example design and through participants’ imitation of generated images. Fixation could also shift from the original example onto AI-generated imagery.

  • Fixation from prompts: Prompts using brief keywords or example-related terms often produced conceptually similar AI images, providing an initial stimulus for fixation.44% of AI-generated images were conceptually similar to the example design; 66.6% of prompts included robot-related or brief-derived terms.
  • Fixation from prompts: Participants frequently created prompts from the design brief or example robot, narrowing their exploration toward related concepts.Participants also created 39 prompts without brief words or robot-related phrases, showing some attempts to explore beyond the initial design.
  • Fixation displacement: A strategy for overcoming fixation was generating prompts beyond the brief and example, then abandoning intermediate ideas to explore substantially different designs.The paper presents this as a quality that may be needed from AI systems intended to support designers in avoiding fixation.
  • AI-generated imagery: The design fixation score of each sketch was moderately positively correlated with the score of the immediately preceding AI images (ρ= 0.56).This quantitative pattern supports a relationship between fixation-related features in AI images and fixation in subsequent sketches.
  • AI-generated imagery: Participants often imitated AI-generated images, producing sketches similar to both the generated images and the original example design.The AI images and sketches were qualitatively similar, and interview data indicated that participants sometimes copied what they observed and then modified it slightly.
  • Fixation displacement: Fixation displacement occurred when participants moved away from the original robot design but became fixated on AI-generated imagery instead.One participant used the unrelated prompt “goddess,” then produced and iterated on sketches of a woman’s face resembling the returned images.

5 DISCUSSION

The study found that AI-supported ideation increased fixation and reduced several divergent-thinking outcomes, with fixation emerging through prompting and responses to generated images. The discussion identifies design strategies and scope limitations for AI-supported ideation.

  • AI-generated images induced higher design fixation, while fluency, variety, and originality were lower than in the baseline condition.
  • Prompt creation became a potential fixation point because participants reused brief keywords, producing images with features similar to the initial example.Participants tried to avoid copying the initial example during sketching, but the fixation measure still reflected replicated features.
  • High-fidelity, visually rich outputs might amplify conformity and displace fixation from the initial example onto AI-generated images.The authors connect this possibility to prior findings that complete and strong examples can cause fixation.
  • Participants’ sketches shared features with AI-generated images, indicating imitation or direct copying across different ideation strategies.Participants gravitated toward generated features whether they ideated on the fly or considered multiple ideas first.
  • With equal task time, interacting with AI or image search reduced sketching time and was associated with lower fluency, creating a trade-off between seeking inspiration and producing ideas.Participants who spent more time refining prompts and interacting with AI had worse ideation performance.
  • AI co-ideation tools could use timely warnings, varied abstraction, alternative directions, and visual-analogy scaffolding to encourage ideation beyond copying.The proposed strategies include partially completed or blurred outputs and support for recognizing abstract correspondences.
  • The findings describe somewhat naïve tool use because generative AI interaction paradigms were still emerging and users were still learning to use them.Results may evolve as users become accustomed to generative AI tools and incorporate them into practice.
  • The 20-minute task limits the findings to short-term rapid ideation, and the study makes no claims about how expertise affects the observed effects.Longer reflection, incubation, sketch iteration, and professional experience could produce different results.

6 CONCLUSION

The study provides empirical evidence that AI image generators can hinder novice designers’ divergent ideation under the tested conditions. It argues that co-ideation tools should support effective ideation behaviours rather than only generate stimuli.

  • The study contributes empirical evidence to debates about whether generative AI can augment human creativity.
  • For novice designers, AI image inspiration produced higher fixation and lower fluency, variety, and originality than image search or no inspiration support.
  • Fixation may arise through the brief and example shaping the prompt, the system translating the prompt into images, and participants using those images for ideas.
  • Under fixed time for visual ideation, the authors suggest spending time sketching may be better than seeking inspiration through AI.
  • Co-ideation systems should encourage effective ideation behaviours and reduce obstacles such as design fixation, not only generate stimuli.

A THE MEAN SCORES AND STANDARD ERRORS OF THE NASA TASK LOAD INDEX (NASA-TLX) SCALES

Table A.1 reports mean scores and standard errors for six NASA-TLX scales across the No support, Image Search, and GenAI conditions.

  • Table A.1 presents mean scores and standard errors for mental, physical, temporal, performance, effort, and frustration demand across three conditions.
Loading 2403.11164v1…