The Effects of Generative AI on Design Fixation and Divergent Thinking

Among the factors that hinder designers’ creativity, “design fixation” is one of the most well-studied phenomena in creativity and design research. It is identified as the unconscious adherence to a set of pre-known ideas or knowledge that restricts the ideation space [30, 76]. When a person experiences design fixation, they tend to adhere to pre-conceived ideas and concepts, limiting exploration of the design space during ideation [30, 40, 47]. Design fixation narrows designers’ ability to explore the creative space between abstract ideas and potential solutions [30, 64]. Previous findings show that this is reflected heavily in their design outcomes and restrains designers from maximising their creative potential, resulting in unoriginal outputs [30].

Design fixation has been studied extensively across different fields [61], including cognitive science [9], design [5, 34], education [28], mechanical engineering [67, 74], and psychology [4, 57]. These studies have collectively shown that design fixation is more likely to occur when designers are exposed to example solutions for design tasks [30]. It has also been demonstrated that the modality, degree of abstraction of the inspiration (i.e. the fidelity), and the designer’s level of expertise [1] can affect fixation intensity when exposed to external stimuli.

Fixation has typically been studied through quantitative experimental approaches in which participants are asked to solve a design problem, either with or without an example (external stimuli) [64]. For instance, Jansson and Smith’s classic design fixation work [30] reported four experiments. These experiments divided participants into two groups: a treatment group (fixation group), who were given a design problem along with an example solution, and a control group, who were given the same problem with no examples to work from. They hypothesised that showing an example design would restrict the ideas of the treatment group because it would make the participants fixate on the given example. Jansson and Smith [30] found that even though both groups produced a similar number of designs, ideas in the fixation groups were more similar to the example. In a subsequent experiment, the researchers found that the flexibility and originality of the designs were limited in the fixation group and concluded that creative performance may be inhibited when an example induces design fixation. Since then, several studies have been conducted replicating or amending the method and examples [38, 64].

When looking at design fixation, it is important to distinguish different types of fixation effects [18]. Youmans and Arciszewski [76] identify three such effects. The first is unconscious adherence [76] to past designs without realizing. An example of this is copying the features of an example (even if the features are inappropriate to the task) [18, 30].The second is conscious blocking [76], where new ideas are actively but perhaps momentarily dismissed. In this situation, a designer is aware of alternative creative paths but chooses to disregard them, perhaps due to a commitment to a current project’s direction or a bias towards familiar solutions. The third type of fixation effect is intentional resistance, a deliberate decision against exploring new concepts. For instance, design companies engaged in research and development often prefer to explore solutions that fall within their well-established expertise, a tendency known as local search bias [18, 50].

Apart from trying to understand its causes, researchers have explored various strategies to overcome design fixation [64]. Such strategies include incorporating physical prototyping in the ideation activities [67], triggering frequent reminders for participants to consider all available options in a timely manner during an ideation task [42, 76], utilising design thinking and lateral thinking methods [6] such as de Bono’s six thinking hats [2, 20], having short breaks or “incubation periods” during the task [58, 73], using computer-aided design and intelligent agents [19, 27], and incorporating design by analogy [12].

Even though there is a large body of work on design fixation exploring the effects of external stimuli on creative tasks [1], studies centred around design fixation are limited within the field of HCI. Among these few studies, HCI researchers have started to examine the potential of using AI image generators as tools for supporting creativity [27, 39]. Thus, in this study, we adapt experimental methods from mechanical engineering and design research, where design fixation is framed as unconscious adherence and is measured by the degree to which participants directly copy features from an example stimulus. We aim to understand the influence of AI image generators on design fixation and divergent thinking, adding new empirical evidence to the HCI literature.

Since the early 1990s, designers have envisioned a future with intelligent design and creative aids [24]. With recent advances in Generative AI, this vision is becoming a reality. Generative AI systems can create new, plausible media [49] to aid individuals in creative tasks [31]. Generative AI models are trained on large data sets and can enable people to generate content such as images, text, audio, or video quickly and easily [35]. Currently, Generative AI tools enable users to create diverse artefacts by providing instructions in natural language called “prompts”. Generative AI systems can also synthesise diverse concepts and generate unpredictable ideas. In the case of AI image generators — the focus of our study — the output comes from the latent space of a deep learning model, arising from an iterated diffusion process that involves the model arranging pixels into a composition that makes sense to humans [66]. Because of process randomness, different results can be obtained based on the same prompt, with entirely new images each time. This differs from conventional image searches, where the search is performed by entering a query into a database to retrieve images that the search engine considers relevant. Another difference is that whereas long and specific queries might be too restrictive for an image search engine, they can benefit AI image generators.

Previous works have explored the roles that generative AI can play in the creative process [29]. For instance, AI can generate content entirely by itself with instructions from the user, or it can act as a creativity support tool, augmenting the user’s creativity [43]. AI text generators can be used as a tool to define specific problems to solve and promote convergent and divergent thinking [72] and have the potential to be used as a co-creative assistant for a designer [19, 54]. Professionals in creative industries claim that AI could be a promising tool to gather inspiration [3].

With the growing interest in AI, HCI researchers have also started to explore ways of using AI as a creativity support tool [16, 31]. Among these explorations, a growing stream of literature focuses on using generative AI to access inspiration and mitigate design fixation. Researchers speculate that generative AI will become a potential solution for inspiring designers [37, 51, 55] due to the ability of AI generators to create abstract and diverse stimuli [32].

One of the early examples in HCI for incorporating AI to mitigate design fixation was the Creative Sketching Partner (CSP) [19, 32], an AI-based creative assistant that generates inspiration for creative tasks. Through multiple studies, Davis et al.  [19] suggest that the CSP helped participants in ideation and in overcoming design fixation. Hoggenmueller et al. have also explored how generative text-to-image tools can support overcoming design fixation experienced in the field of Human-Robot Interaction [27]. They conducted a first-person design exploration and reflection using “CreativeAI Postcards” inspired by Lupi and Posavec’s “Dear Data book” method to ideate and visualize robotic artefacts. They noted that AI-generated images have the potential to inspire new robot aesthetics and functionality and also claimed that the designer’s AI-co-creativity can help to eliminate biases and expand limited imagination. In a different case, Lewis [39] reflects that a digital assistance tool like “ChatGPT” helped her by acting as an art teacher and providing instructions. Lewis points out that it is challenging to distinguish between inspiration and copying when utilizing generative AI and reflects on concerns such as “transparency of attribution”, “ethical considerations”, and the clarity of the “creation process”. Rafner et al. [48] conducted an in-the-wild study to examine the effects of AI-assisted image generation on creative problem-solving tasks, aiming to investigate the effects of generative AI on problem identification and problem construction. They developed a human-AI co-creative technology that combines a GAN and stable diffusion model to support AI-assisted image generation. They found that this intervention enabled participants to facilitate idea expansion and prompt engineering, suggesting that AI can “aid users in generating new ideas and refining their initial problem representations” [48].

As the domain of AI-powered creativity support is still in its infancy, the available literature provides only a nascent understanding of the effect of AI on creativity and design fixation. Our work extends the literature by using established techniques from design fixation research to better understand how AI image generators affect design fixation during a visual design task.

The experimental task consisted of a visual ideation activity in which participants were asked to devise as many ideas as possible for a new chatbot avatar by sketching them on paper. The written design brief given to participants was:

This written design brief included an example of an avatar with the figure caption "Example chatbot avatar (for reference only)". The example avatar is shown in Figure 1.

Further, we provided verbal instructions for the participants, asking them to produce as many different ideas as they could during the experiment. For participants in the Image search and Gen AI conditions, we additionally informed them that they could use the digital tool (either Google Image Search or Midjourney, depending on the condition) to gather inspiration for their work. The full study protocol can be found in supplementary material.

Similar to previous work [30], we started the task by showing participants an example avatar to induce design fixation. We drew inspiration from Ward’s creature invention task [36, 70], which asked participants to imagine and create animals that lived on a different planet. The authors of this paper created the example chatbot avatar after several design iterations. We created the avatar so that it had 14 salient features, which we used to quantitatively assess design fixation (see Figure 1). We considered the presence of these features in participants’ ideas to be evidence of design fixation, following standard practice in the literature [30]. In the experimental task, participants were given 20 minutes to sketch their ideas for addressing the brief. We chose this time limit because it is the median time given to participants in previous design fixation studies [64] and because we aimed to cap each experimental session at one hour to avoid fatigue. We provided participants with pencils, pens, felt pens, and coloured pencils, along with blank A4 sheets to sketch their ideas. A timer was placed outside their peripheral view for them to keep track of time.

The experiment included a single between-participants independent variable—the Inspiration Stimulus available during the task—with three levels:

We assessed participants’ creative output using four standard measures from the design fixation literature: design fixation, fluency, variety, and originality, which we describe as follows:

Design fixation is the unintentional conformity towards existing ideas or concepts that limits exploration of the ideation space [30, 76]. Researchers use the degree of copying as a method to quantify design fixation [30, 45]. Therefore, we operationalise design fixation as an objective property of each sketch based on the presence or absence of features available in the example. Following the approach used in design fixation literature [45], two raters blind to the experiment’s aims counted the presence of features from the example avatar in the sketches created by the participants. We validated the ratings by computing the inter-rater reliability and computed the design fixation score (DFS) as follows:

Fluency refers to the number of ideas produced by the participants  [25, 62]. We operationalise it by counting the number of sketches produced by each participant within the available time (20 minutes).

Variety measures the coverage of the solution space explored during the idea-generation process [53]. It aims to capture the extent of the design space covered during ideation. If the majority of ideas are similar, it indicates less variety. To compute variety, we assigned a numerical identifier to all the sketches (N=277), imported them into a Miro⁷ (an online collaborative whiteboard), and displayed them in randomised order. Two raters (blind to the conditions) iteratively and inductively grouped similar sketches into mutually exclusive clusters. This activity considered several factors: appearance, embodiment, appendages, shape, and accessories. The process resulted in 83 clusters. Each participant received a Variety score based on the number of clusters their sketches were classified into. We subtract 1 from the number of clusters so that if all of a participant’s sketches belong to the same cluster, their score is 0, and if they have sketches in every cluster, their score is 1.

Originality (also called Novelty [23, 53]) refers to the uniqueness of a particular sketch within the total pool of sketches made by participants [25, 30]. It measures how unusual and unexpected a given idea is. Intuitively, the more people have the same idea, the less original it is. We computed an idea’s originality by counting the number of other participants who had an idea belonging to the same cluster, dividing it by the total number of other participants, and computing its complement to 1 (to normalise the value between 0 and 1). In other words, it is the proportion of other participants who did not have the same idea. This score is 0 when every participant had an idea in the same cluster and 1 if only a single participant had an idea in that cluster.

We recruited 60 participants through digital student notice boards, mailing lists of university student clubs, and word of mouth. Participants expressed their interest through a digital signup form. Participants self-described their prior experience in visual design (measured in years/months). We did not specify this experience should only be professional design experience. We screened participants based on our eligibility criteria and invited those who were 18 years or older and had experience in visual design via email. Further, to avoid dependent relationships, we ensured that none of the participants had a direct connection with the primary researcher running the study. Participants had a mean age of 25.8 years (18–49, SD = 5.4). They included undergraduate, master’s and PhD students from diverse domains such as arts, business, computer science & IT, design, engineering, and science. Each condition had an equal number of participants and was gender-balanced, with 10 women and 10 men per condition (gender was self-described by participants).

Participants booked a time to participate individually based on their availability. The study was carried out in a quiet research laboratory. Upon arrival, participants read a plain language statement describing the study and consented to participate (Figure 3-1).

The experiment had four stages: pre-study questionnaire, main experimental task, post-study questionnaire, and semi-structured interview. Each session lasted 45–60 min in total. In the pre-study questionnaire, we collected participants’ basic demographic information, their experience with similar design tasks (measured in years/months), and their familiarity with AI image generators (a yes/no question, and participants were asked to list any systems they had used if they answered yes). (Figure 3-2). The main objective of this questionnaire was to understand and control for any variables that might confound the results. After completing the questionnaire, the participants were randomly assigned to one of the three conditions and were assigned a unique ID generated by the computer (3 random digits) (Figure 3-3).

In the main experimental task (Figure 3, steps 3-7), participants in all conditions received the same design brief, which asked them to design an avatar for a chatbot in 20 minutes, as described in Section 3.1. We started by allowing participants to familiarise themselves with the available materials. Then, the participants assigned to the image search and AI-supported groups received an introduction to the tool they would use during the design task (Figure 3-5). These tools were available for them to use on an Apple MacBook M1 Pro laptop. The tool introduction included a video tutorial created by the research team. This video tutorial explained how to use the tool. After the video tutorial, we allowed participants to ask questions and clarify any doubts.

We provided task instructions to participants both verbally and as a written brief. The written brief included an example of a chatbot avatar, which served as a stimulus to induce design fixation (Figure 3-6). Participants were given 20 minutes to complete the design task (Figure 3-7). We limited the design task to 20 minutes to minimise the possibility of fatigue and because previous studies considered it an ideal duration for maintaining focus for producing ideas with both quality and quantity [64, 74]. Once participants indicated they were ready to start, the researcher started the screen recording with participants’ consent (in Image Search and GenAI conditions), switched on the timer and left them alone to work in the room, allowing them to work independently.

After the design task, the researcher entered the room and asked the participant to fill in the post-study questionnaire (Figure 3-8). As the post-study questionnaire, we administered the NASA-TLX [26] to ensure that all conditions induced an equivalent workload. To analyze the NASA-TLX, we used a one-way ANOVA; the effect of the independent variable "condition" on the NASA-TLX overall score was not statistically significant (F(2, 57) = 1, p = 0.37). Therefore, we did not conduct post-hoc tests.

Then, the researcher conducted a semi-structured interview. Each semi-structured interview lasted 15–20 min. Through the semi-structured interview, we aimed to get insights into the participant’s background and their past experience in creating logos and avatars. We also probed for possible feelings of design fixation during the experiment and how it was affected by their previous knowledge, experience and process. In addition, we asked questions to understand how the stimuli (or lack thereof) affected their ideation process. To conclude the study, we debriefed the participants about the purpose of the research. We thanked each participant with a $20 gift voucher.

We scanned all the sketches participants created and assigned them a unique identifier. Two independent evaluators rated the sketches to compute the design fixation score, variety, and originality measures. These evaluators were researchers from the human-computer interaction domain with experience in teaching and evaluating design.

We extracted all prompts and images from the Midjourney gallery where the logs were saved (not visible to the participants), compiled the sketches and arranged them in the sequence in which they were created as a visual sequence board. Underneath the AI-generated images, we added the sketches of the participants (Figure 4).

We used a mixed-method approach for our analysis. For quantitative analysis of design fixation and divergent thinking, we built Bayesian statistical models to quantify relationships between our dependent and independent variables (see section 4.1). We employ Bayesian statistical methods to analyze our results, opting for this approach due to its added flexibility, capability to quantify uncertainty, better handling of small samples, and greater potential for future extensibility. For a comprehensive rationale advocating the use of Bayesian methods over traditional frequentist statistics in the field of Human-Computer Interaction (HCI), see Kay et al. [33]. Readers who may not be familiar with these methods can find a beginner-friendly introduction in McElreath [44] and can see examples of their practical application in HCI in Schmettow [52]. In this manner, we shift the focus away from p-values and dichotomous significance testing, directing our discussion towards causal modelling and parameter estimation.

For qualitative analysis of participants’ interview data, we used Braun and Clarke’s 6-phase approach to reflexive thematic analysis [7, 60]. The analysis was inductive, i.e. data-driven, based on transcripts of the interviews. Each phase of the analysis was progressed using NVivo12⁸ for coding procedures, theme development and naming. The analysis aimed to understand potential causes of design fixation during the experiment and participants’ approaches to creating sketches in each condition. In this paper’s findings, we use interview quotes to illustrate participants’ approaches to prompt creation and their stated approaches to ideation based on AI images. This enables us to probe plausible explanations for observed differences between experimental conditions and explore why particular kinds of sketches were created in response to AI-generated images.

We summarise our theoretical claims as a directed acyclical graph (DAG) in Figure 5. We argue that the Inspiration Stimulus affects users’ Design fixation, Fluency, Variety, and Originality. The choice of inspiration stimulus affects how much time is spent on the sketching task (as opposed to seeking inspiration), which, in turn, affects the number of sketches produced (fluency). Higher fluency is also likely to lead to higher variety—as producing more sketches also increases the likelihood that they will cover more ground during ideation. A greater variety of sketches, in turn, will likely lead to more original ideas.

We used the brms package [8] to fit our models. This package facilitates the implementation of Bayesian multilevel models in R, leveraging the Stan probabilistic programming language [11]. To ensure the reliability of our Bayesian Markov Chain Monte Carlo (MCMC) sampling process, we assessed convergence and stability through two metrics: R-hat, which ideally should be less than 1.01 [65], and the Effective Sample Size (ESS), which should ideally exceed 1000 [8]. All of our model estimates met these criteria. We built our models based on the original count data in the direct measurements but report normalised values as described in Section 3.1 in our plots for easier comparisons with future work.

In our reporting of model results, we present the posterior means of parameter estimates, their corresponding standard deviations, and the boundaries of the 89% compatibility interval, often referred to as the credible interval. The choice of an 89% compatibility interval aligns with the recommendation by McElreath [44] to mitigate potential confusion with the frequentist 95% confidence interval, as the two intervals have distinct interpretations. The compatibility interval specifies the range of values within which there is an 89% probability that the true value lies. We report hypothesis test results using Bayes Factors, which compares the likelihood of the observed data under the proposed model over the null. We interpret these values following Wagenmakers et al. [69], considering values above one as supporting a given hypothesis, values under 3 offering anecdotal evidence; under 10, substantial evidence; under 30, strong evidence; under 100, very strong evidence; and above 100, extreme evidence. We note that p-values are not used in Bayesian statistics, and no claims about “statistical significance” should be derived from our results.

To model Design Fixation, we consider the number of salient features in participants’ sketches also found in the example avatar provided at the beginning of the experiment. We model this data as a binomial distribution with N = 14 (the maximum number of features) and a probit link. We use weakly informative, regularising priors for the model parameters (drawn from a normal distribution with mean zero and standard deviation of 2). We model the random effects of participants and images as being drawn from a normal distribution with mean zero and standard deviation computed from the data through partial pooling.

The model suggests that the effect of GenAI has a 100% probability of leading to higher Design Fixation (mean = .32, 89% CI [.11, .55]), and a Bayes Factor of 124 suggests extreme support for the hypothesis that inspiration from GenAI leads to higher Design Fixation than the baseline. The effect of Image Search on Design Fixation was also detrimental, but less so (mean = .27, 89% CI [.05, .49]), with a 98% probability of this effect leading to higher Design Fixation. A Bayes Factor of 42.48 suggests very strong evidence for the hypothesis of a higher Design Fixation than the No support baseline. In summary, our model suggests that, on average, both stimuli led to more features in common with the example avatar, and GenAIled to even more design fixation thanImage Search.

To model Fluency, we consider the number of sketches produced by each participant. Our causal model considers two effects of the stimulus on Fluency: a direct effect and an effect mediated by Time on task (ToT). We model these effects through two models, with and without Offset(log(ToT)) as a covariate. This approach was taken to account for varying time on task as participants of both GenAI and Image search utilised different times to sketch. In both cases, we model the expected value for each response as a negative binomial distribution with a log link. This models the sketch count based on the mean and the shape parameter, both of which depend on the inspiration stimulus. We opted for this model instead of a Poisson model due to its ability to model overdispersion. We use weakly informative, regularising priors for the model parameters (drawn from a normal distribution with mean zero and standard deviation of 2).

In the total effects model, both Image Search and GenAI demonstrate detrimental effects on Fluency. Specifically, for GenAI, there is a notable negative impact on Fluency with an estimate of -.21 (89% CI [-.7, .28]) and a Bayes factor of 3.20, suggesting a 76% posterior probability of a negative effect. This is a substantial indication of its negative total effect on Fluency. The effect of Image Search is more pronounced (mean = -.49, 89% CI [-.9, -.10]) with a Bayes factor of 46.62, indicating a 98% posterior probability of a negative effect, strongly supporting its detrimental influence on Fluency.

In the direct effects model, which accounts for the time-on-task, the impact of Image Search on Fluency is minimal (mean = -0.22, 89% CI [-0.64, 0.19]) with a Bayes factor of 4.20, indicating an 81% posterior probability of a negative effect. In contrast, GenAI shows a relatively neutral direct effect on Fluency (mean = .10, 89% CI [-.41, .60]) with a Bayes factor of .61, implying only a 38% posterior probability of a negative effect.

In summary, neither Image SearchnorGenAI enhanced Fluency compared to the baseline, with both generally resulting in lower Fluency. The effect of Image Search on Fluency is minimal when controlling for total output time, indicating a less direct impact. However, GenAI does not exhibit a considerable direct negative effect on Fluency, highlighting its influence is not strongly dependent on the total time available for sketching.

To model Variety, we consider the number of clusters a participant’s sketches belong to minus one to account for the fact that variety only begins with the second sketch. Our causal model considers two effects of the stimulus on Variety: a direct effect and an effect mediated by Fluency. We model these effects through two models, with and without Fluency as a covariate. In both cases, the expected value for each response is based on a negative binomial model with a log link. We use weakly informative, regularising priors for the model parameters (drawn from a normal distribution with mean zero and standard deviation of 2 for coefficients and a gamma distribution with parameters set to 0.01 for the shape).

The total effects model shows a detrimental effect of both stimuli on Variety. The model suggests that the effect of GenAI has only an 86% probability of being negative (mean = -.29, 89% CI [-.75, .16]), and a Bayes Factor of 5.93 provides substantial evidence for the hypothesis that it yields a negative total effect on the variety of output. The effect of Image Search was even more negative (mean = -.40, [-.87, -.07]), with a Bayes Factor of 11.38 providing strong support for the hypothesis that it has a negative effect.

The model including Fluency as a covariate, models the direct effect of the stimulus on the Variety of the output. Comparing the two models, we see that after accounting for the number of sketches that the participant produced, Google Image Search did not have much of an effect on Variety (mean = .02 [-.39, .43]. However, GenAI still had a negative effect (mean = -.15 [-55, .24]), with a Bayes Factor of 2.64 suggesting anecdotal evidence for this effect being negative. In summary, neither Image SearchnorGenAI provided meaningful support over the baseline in terms of enhancing the variety of the output, yielding, on average, lower variety than the baseline. The effect of Image Searchwas fully mediated byFluency, butGenAIalso had an additional negative direct effect onVariety.

To model Originality, we consider the number of other participants with sketches in the same cluster as each sketch. As in the case of Variety, our causal model considers two effects of the stimulus on Originality: a direct effect and an effect mediated by Variety. We model these effects through two models, with and without Variety as a covariate.

The expected value for each response is based on a linear regression model. This models the originality score based on the inspiration stimulus (and variety score in the direct effects model), as well as a random effect of the participant. We use weakly informative, regularising priors for the model parameters (drawn from a normal distribution with mean zero and standard deviation of 2 for coefficients). We modelled our random effects as being drawn from a normal distribution with mean zero and standard deviation computed from the data through partial pooling.

The total effects model suggests that both stimuli had small negative effects on Originality. The model suggests that the effect of GenAI has a 97% probability of being negative (mean = -.03, 89% CI [-.06, .00]), and a Bayes Factor of 35. provides extreme evidence for the hypothesis that it yields a negative effect on the variety of output. The effect of Image Search was slightly less negative but also rather small (mean = -0.01, [-.04, .01]), but a Bayes Factor of 3.9 provides only anecdotal evidence against the hypothesis that it has a positive effect. Adding Variety did not change the model in any meaningful way, suggesting that Variety does not mediate an effect on Originality. In summary, neither Image SearchnorGenAIprovided a considerable aid in terms of developingOriginality of the output, offering, on average, lower originality than the baseline, but these effects were negligible.

The results from our statistical models suggest that support from Generative AI led to higher design fixation. To understand why this occurred, this section draws on our interview data, the prompts created by participants, the AI-generated images, and the participants’ sketches. We first explore the content of participants’ prompts as one potential cause of design fixation. This encapsulates how participants claimed to develop the prompts and how they were influenced by the design brief or the example design. Next, we quantitatively explore the similarity between participants’ sketches and the AI images used to inform that sketch in terms of design fixation. We then discuss the types of AI-generated images returned by participants’ prompts and the sketches created based on them, using a case-study-based approach to illustrate our claims.

Overall, our analysis indicates that participants frequently relied on prompts containing keywords copied directly from the design brief or used prompts inspired by the example design. These prompts resulted in AI-generated images that were conceptually similar to the example design in 44% of cases and which frequently contained fixating features that were present in the example design. Further, while not all sketches exhibit high similarity to the example we provided, ideating based on AI images can lead to fixation displacement, where participants simply fixate on the images generated by the AI and copy what they see. This can occur irrespective of whether the participant imitates the example design or whether they attempt to explore other areas of the conceptual space.

Through our study, we identified that the prompt acted as a potential source of design fixation. While participants used different strategies to devise their prompts (Figure15-A), the results suggest that most of them used keywords from the brief (Figure15-A₁) and built upon the idea of a ‘robot’ from the initial example (Figure15-A₂) when devising their prompts. Participants claimed that they tried to avoid copying the initial example later during sketching (Figure15-B), but this was not reflected in the data given that the high design fixation score was calculated based on the ratio of replicated features from the initial example. Therefore, we assume that poor prompt design led to the generated images sharing similar features with the example. This suggests that design fixation happened when the participants created the prompts (Figure15-A). Participants tended to produce prompts that were semantically similar to the words given in the design brief and examples, while exhibiting strategies of repeating the same steps when creating prompts. Prior work indicates that participants can become fixated on exposed examples [10, 30]. Further, studies have shown that participants tend to fixate on self-generated ideas and concepts compared to initial examples [38]. Our study aligns with these findings, suggesting that some participants were fixated on the design brief, the example avatar, or self-generated ideas when determining keywords for the prompt.

Thus, paying attention to different prompting strategies might help mitigate this first potential occurrence of fixation when co-ideating with AI. Youmans et al. [76] summarise different strategies to mitigate design fixation based on the cause of occurrence. One strategy to snap participants out of fixation is to have timely warnings to consider alternative options [76]. Therefore, creativity support systems based on generative AI could not only turn prompts into images but also scaffold users’ abilities to craft better prompts for ideation. AI systems could support users in creatively interpreting the brief, push users’ thinking into alternative directions, or mix arbitrary ideas into the prompts. This functionality could be enabled through other generative AI techniques, such as large language models.

Cheng et al. [13] have found that showing low-fidelity, abstract, and partially completed ideas led participants to become more divergent in their thinking and reduce fixation. Users of AI systems should consider using prompts that generate low-fidelity abstract or partial images when interacting with AI because images with these qualities might alleviate design fixation and encourage divergent thinking in an ideation task. Having a predetermined prompt structure or template that describes ways to make the images more abstract and less refined might lower the risk of fixation.

The images generated by the AI system in this study were high in fidelity, visual detail, and quality; they appeared to be rich in shape, form, texture, colour, composition, and visual expressiveness (see Figure 11-14). Though this showcases impressive functionality, it might have amplified conformity towards the generated image, causing fixation displacement. This aligns with prior studies, which have shown that complete and strong examples carry the potential to cause fixation [10, 13, 15, 59]. Previous work has found that introducing some incubation time can help dissolve concrete exemplars into more abstract concepts, lowering fixation [58, 73] and supporting the emergence of novel ideas [21, 56]. Though this was not possible in our study, given the short time available for the task, it is a process that users of AI systems can incorporate into their ideation process. Further, when developing generative AI to support ideation, it may be useful to introduce mechanisms to lower the fidelity and the richness in detail of the output. Another direction is to show partially completed or blurred outputs, which might be beneficial for introducing ambiguity and pushing ideas in new directions [13]. Recent works by Davis et al. [19] and Williford et al. [71] provide initial evidence suggesting that these mechanisms might be plausible approaches to be embedded in generative AI.

Through our visual comparisons, we observed similarities between the sketches produced by participants and the AI-generated images, suggesting that participants had imitated and, in some instances, directly copied elements from the images generated by Midjourney. Further, we identified that regardless of whether participants ideated on the fly or considered multiple ideas before creating a sketch, they gravitated towards features of the images which Midjourney generated, leading to fixation. Copying elements from an example is the easiest way to result in fixated outputs [10, 30], and our findings were consistent with this.

To successfully act as sources of inspiration, generative AI tools must encourage more strategies that are more effective than copying. Previous work has shown that techniques like visual analogy—identifying abstract correspondences between the images being generated and the solution being sought—can improve the ideation effectiveness of designers of all levels, including novices [12]. However, Casakin and Goldschmidt highlight that even though novices already have an inherent understanding of how visual analogy works, they must be shown how to do it well and how it can support problem-solving in design activities  [12]. Scaffolding these skills is a promising role for AI-based creativity support tools.

We observed lower fluency in the GenAI and Image search conditions. Because the time given to complete the task was the same in all conditions, there was an inherent trade-off between spending time producing ideas vs. seeking inspiration. Interacting with both the AI image generator and the web image search led to less time spent sketching. These results tally with the findings of Vishwanathan and Linsey [67], who found that though physical prototyping techniques that required more effort led to higher quality ideas in an engineering design task, they also increased design fixation and lowered fluency. They hypothesise that this is due to a “sunk-cost effect”—the higher the effort spent in a given direction, the harder it is to move into a different one. Participants who spent more time refining prompts and interacting with the AI also had worse ideation performance. Users of generative AI systems should be careful and deliberate in their approach when seeking inspiration from external stimuli like AI image generators to mitigate the risk of design fixation. Crilly [17] suggests that empowering designers to recognise and reflect upon fixating episodes might be beneficial in developing a less fixating co-ideation workflow with AI. Further, Neroni and Crilly [46] state that uncovering participants’ fixation tendencies, which they call "demonstrated vulnerability", is an effective approach that can further strengthen participants’ ability to overcome fixation. In summary, when developing generative AI for co-ideation tasks, there is a rich opportunity for designing interactions with intelligent agents that not only generate stimuli but also encourage better ideation behaviours. Triggering timely reminders, suggesting new idea directions, preempting fixation, varying the abstraction of the visual outcomes, and facilitating visual analogical reasoning are all promising directions for future work.

We acknowledge several limitations in our study. First, at the time of writing, generative AI tools are still nascent technologies. Interaction paradigms are emerging, and users are still learning to leverage their potential. As such, our results describe a picture of somewhat naïve use of these tools. It will be interesting to see how these results evolve as users become accustomed to generative AI tools and incorporate them into their practice.

Next, in this study, we gave all groups of participants the same amount of time to complete the task, but we observed lower fluency in the GenAI condition. We note that when the experiment took place, the AI system did not produce results instantaneously, which potentially delayed participants in that condition. However, we also note that participants in the Image Search condition, who did get their results instantaneously, also exhibited lower fluency than the baseline. This could be due to the exposure to a large number of images with endless scrolling, which added another layer of decision making to pick the ideas that suit them. Therefore, in a real-world setting, it is important to consider the trade-off between spending time on the task (e.g. by sketching) vs. seeking inspiration (e.g. by interacting with a creativity-support tool).

We acknowledge that because we limited the task time to 20 minutes based on previous work, we restrict the scope of our insights to short-term usage of these tools in a rapid ideation task. In the real world, people may spend longer reflecting on the outputs of AI, and incubation time along with iteration of sketched ideas may produce results different from those of our experiment. Further, we screened participants for prior skills in visual design, but few had professional industry experience. Though our sample was balanced across conditions, we make no claims about how these effects might be affected by expertise.Therefore, with this study, we can only provide initial insights into how a novice designer might approach a design task, and to generalize these claims, we need further investigation.

In this study, we operationalised design fixation by looking for a restricted set of salient features from the example in participants’ sketches. Though the choice of focusing on denotative elements of the design aims to facilitate operationalization, we acknowledge that there are also connotative aspects that were left outside the scope of our analysis, including art style, emotional expression, and cultural references. Finally, our study only evaluated the potential of generative AI tools for ideation support through the specific example of image generators in a visual ideation task. It remains to be seen how these effects translate to other modalities, such as text, video, audio, and music generation.