BI-CST: Behavioral Science-based Creativity Support Tool for Overcoming Design Fixation.

’Design Fixation’ refers to the uncritical adherence to pre-existing ideas, impeding the exploration of innovative conceptual designs  [15]. This phenomenon obstructs the incorporation of ideas already present during the problem-solving phase of design. Fixation in the design process can involve both conscious and unconscious adherence  [30] and tends to be more pronounced in the early stages of the design process, specifically during ’early ideation’ or ’idea generation’  [6]. This tendency is referred to as confirmation bias in cognitive psychology  [23].

Various research endeavors have been undertaken to mitigate Design Fixation. At the strategy level, there have been suggestions on techniques such as redefining problems and using clues or hints to provoke new ideas  [26]. At the experimental level, diverse intervention methods have been proposed, including providing partial photographs  [3], text stimuli to stimulate novel thinking  [11], or multi-modal triggers such as video and animation  [25]. However, precedent research's stimuli are limited to photographs or keywords, which cannot provide the context of human behavior, which is crucial in user-centered design  [1].

Computer-based Creativity Support Tools (CST) have been widely researched to aid humans’ creative problem-solving in the HCI field. They have been studied not only for generating algorithms or facilitating style transformations  [4] but also for aiding information-based ideation for shifting perspectives  [18], idea generation  [21], providing metaphor creation  [8], and co-creating design work  [10, 20]. Since the advent of generative AI, such as Chat GPT and Dall-E, research on applying generative AI to generate images has increased, demonstrating that AI can assist in creative ideation  [2, 17]. However, there is limited research on trained Large Language Models (LLMs) for providing ’contextual information on human behavior’ as stimuli.

Behavioral science, a subfield of psychology, aims to understand and predict people's behaviors through scientific methods ranging from casual observation of daily life to systematic observation to minimize the effects of biases  [22, 29]. Its research includes behavioral decision theory, which shows people's heuristics and biases in judgment  [9], and builds nudges that gently alter people's decision-making processes  [27]. As this field targets problem-solving by understanding people's psychology, efforts have been made to apply theories of behavioral science in the HCI field to change users’ behavior [13, 14, 19]. However, there has been limited research on incorporating behavioral science into Creativity Support to aid design processes. We assume that the scientific causal and systematic observational results about human behavior from behavioral science can serve as a ’nudge’ to assist users in gaining more diverse perspectives.

Our approach is based on the Double-Diamond Model of Design  [5], which categorizes the design process into four phases: ’discover’, ’define’, ’develop’, and ’deliver’. These phases represent a cycle of diverging and converging activities essential for creative output. BI-CST aims to support the divergence process by providing insights from behavioral science research, thereby helping to overcome design fixation and enhance the quality of creative outputs.

A visual representation of the Double-Diamond Model of Design. The diagram shows how BI-CST enhances the divergence phase by introducing fresh perspectives from behavioral science that are unfamiliar to the users, aiding in broadening their conceptual scope.

This composite image comprises two main sections. On the left, the System Architecture of BI-CST is depicted, showing how it processes (A) a problem statement and (B) initial user ideas to (C) generate and select theories and examples using GPT-4, ensuring they offer new perspectives. On the right, the BI-CST User Interface is shown. It displays the theories and examples chosen by GPT-4 that relate to the problem but differ from initial ideas, fostering creative exploration. This interface supports user interaction, allowing for follow-up questions to deepen understanding of the concepts, highlighted in panel (D).

1. Problem Statement and Initial Idea: In BI-CST, the ’Problem Statement’ (A) is initially described and prefilled in the tool for the experiment; however, users have the option to edit it if they are utilizing the tool for their own purposes. Subsequently, users can input their ’initial idea’ (B) into the chat, which should address the problem outlined in the Problem Statement. Additionally, users can interactively pose follow-up questions to the AI via the chat box (D).

2. AI's Advice: After users input their initial idea, BI-CST presents one of the behavioral science theories via a chat bubble (C) that addresses the problem while intentionally offering an approach different from the user's initial idea. Additionally, BI-CST provides two examples tailored to the problem statement, which users can explore further by clicking an expansion button.

BI-CST's LLM is trained on a wide range of foundational research within the field of behavioral science, including ’Heuristics and Biases’  [28], ’Nudge’  [27], ’Social Preference’  [7], and ’Choices, Values, and Frames’  [16]. This training aids users’ divergent thinking by offering new perspectives based on a scientific understanding and predictions of human behavior.

Constructed on OpenAI's ChatGPT platform, BI-CST utilizes a large language model (LLM) capable of generating high-quality, contextually relevant behavioral science content. This content is specifically tailored to each user's problem statement and initial idea. The tool's effectiveness stems from its advanced understanding of context and its ability to dynamically generate content that is informed by a comprehensive dataset of behavioral science theories and research.

To enhance its capabilities, BI-CST integrates natural language processing techniques to interpret and analyze user inputs, ensuring that the generated advice not only reflects diverse scholarly content but also aligns with the users’ cognitive processes. This allows BI-CST to propose novel approaches and perspectives that challenge users’ initial assumptions and promote innovative thinking.

To explore the effects of integrating behavioral science theories into design tasks, we will conduct experiments (with IRB approval), where participants are tasked with designing an application alone or with two different types of generative AI tools. We aim to recruit at least fifteen design and engineering students for each of the three experimental conditions, thereby engaging a minimum of 45 participants in total. Each experiment will be conducted in groups and will last approximately 60 minutes, conducted either in-person in a private meeting room or virtually via Zoom, with each session accommodating up to fifteen participants.

Participants will be assigned to one of three conditions:

• Condition A: Design without any Generative-AI-based CST (no suggestions provided).
  
• Condition B: Design with an AI-based CST without behavioral science guidance (open-ended suggestions).
  
• Condition C: Design with a Behaviorally Informed AI CST (BI-CST) with behavioral science guidance (constrained suggestions in Behavioral Science).
  

This figure illustrates the study design for evaluating different ideation conditions. Participants are allocated to one of three conditions: ideating alone (Condition A), ideating with ChatGPT (Condition B), or ideating with BI-CST (Condition C). The process involves a 15-minute solo ideation phase followed by a 30-minute period to develop solutions, during which participants interact with the designated tool based on their assigned condition. Finally, a 15-minute interview is conducted to gather feedback and insights from the participants.

All participants will receive the same problem statement, which addresses the conflicting needs of stakeholders: “Create an app for family self-tracking of health information that facilitates both parents and children in self-tracking and reviewing health information together.” (Pina et al., 2020). Initially, participants will be given 15 minutes to ideate, with or without the assistance of a generative AI tool (ChatGPT) or a Behaviorally Informed Creativity Support Tool (BI-CST), depending on their assigned condition. Following this ideation phase, each participant will be required to sketch a wireframe and flow for the app using pen and paper, and to add descriptive annotations for 30 minutes. Subsequent to the design phase, 15-minute semi-structured interviews will be conducted to assess participants’ confidence and their comfort with the design process. Lastly, the outputs from the participants will be evaluated by three expert judges, who are both practitioners and educators in the field of design. Their assessments will focus on originality and practicality, rated on a scale of 1 to 5 (Goldschmidt and Smolkov, 2006) [12].

We hypothesize that participants using a non-behaviorally informed Creativity Support Tool (CST) in Condition B are likely to reinforce their existing heuristics, which could increase their design fixation. In contrast, participants in Condition C, who will be provided with behaviorally informed guidance, are expected to adopt new heuristics that help reduce their fixation on initial ideas.

Specifically, we anticipate that the outputs in Condition A and Condition B will likely exhibit medium to high originality but may have low practicality. Conversely, outputs from Condition C are expected to show medium originality but high practicality, due to the application of evidence-based creative support derived from behavioral science research.

Considering that we define creativity as a combination of originality and effectiveness, we expect that behaviorally-informed AI CST can support designers and engineers in creative thinking compared to scenarios without any support or with open-ended suggestions from generic AI.