Generative AI in Creative Writing
MASTER’S DISSERTATION IN HUMAN-COMPUTER INTERACTION DESIGN — 2023
Project Summary
Background
For my Master’s dissertation in Human-Computer Interaction Design, I chose to study the intersection of generative artificial intelligence and creative writing. At the time of my research, I observed that generative AI often appeared as a highly technical solution in search of a problem — so I took an exploratory approach to investigate how the interaction patterns of AI systems affect the creative writing process.
I recruited 10 creative writers as participants in my research study. Following an in-depth interview about their writing process, I asked research participants to complete a series of writing activities using ChatGPT. The activities ranged in specificity and ask of the system. Based on writers’ reactions and ideas during these activities, I created a series of design prompts — high-level, speculative ideas for future AI systems. They ranged in functionality and interaction style. I then used these prompts in participatory design workshops with creative writers, eliciting further reactions and ideas from participants about how different interactions would affect their creative processes.
Based on these activities, I first developed a set of 3 insights to describe how creative writers imagine the use of AI. I then synthesized the data again with a lens towards interaction patterns. I noted real and speculative patterns that may impact the creative process — positively, negatively or in a neutral but meaningful way. I then compared these to the interaction patterns offered by a set of commercially available AI tools, as examples of where future opportunities lie. This artifact is known as a “design space.”
TIMELINE
3 months (July - Sept 2023)
Methods
A significant aspect of the project was to develop tactics for researching human-AI interactions, especially with non-technical individuals. This involved navigating new paradigms and challenges unique to AI technology.
Phase 1: Interviews and Writing Activities
The first phase consisted of conducting in-depth interviews and interactive activities centered around the practice of creative writing. I incorporated writing exercises using ChatGPT, which facilitated deeper discussions and idea generation beyond mere opinions. These activities revealed a diverse array of personal approaches to writing. By categorizing these approaches into core elements such as ideation, writing, and revising, I was able to create a framework for synthesis.
A writer using ChatGPT to generate a story in the style of Vladimir Nabokov, based on a story prompt I provided.
Phase 2: Participatory Design Workshops
In the second phase, I employed participatory design workshops to abstract ideas from the technology. To do this, I created speculative design proposals (shown below) that functioned as probes for discussion. Some of them were intended to seem quite outlandish, to spur discussion, while others were intentionally more concrete. Participants, regardless of their understanding of AI, imagined various ways AI could assist their writing process.
Data Analysis
The project generated a vast amount of data, which I analyzed using a combination of thematic analysis in Airtable and affinity mapping in Miro. This approach allowed me to dynamically link and visualize observations and insights, to keep the research grounded in the raw data. It also enabled easy cross-analysis of diverging or contradictory observations, helping me extract deeper meaning from the data. This was particularly crucial in a small-scale project with limited participants, where relying solely on prominent themes would be inadequate.
An example section of Phase 1 synthesis, in Airtable. All data is dynamically linked back to the individual participant and session, offering a kind of paper trail from observation to insight.
An example section of Phase 2 synthesis, in Miro. Yellow stickies are direct observations or quotes from research; blue stickies are first-level themes; pink stickies are detailed insights; and the green sticky is the name of the high-level category encompassing these insights.
Research Insights
These 3 insights describe how writers imagine possible uses for and possible drawbacks of AI in the writing process.
Generative AI as a Divergent Thinking Partner
For writers, creativity isn’t enhanced by seeing AI generate prose. Instead, workshop participants imagined using AI as a way to connect disparate ideas, create a portal to related concepts, or invoke alternative mediums — always stemming from their own work.
Generative AI for Organizing Complexity
Writers consider their strengths to be creativity, discernment and crafting compelling prose. In contrast, they acknowledge AI’s strengths in identifying patterns, especially in large works. Writers consequently imagined using AI for consistency checks or for suggesting ways to incorporate feedback notes from workshops.
Generative AI as a Distraction
Much of writing is built on subjectivity, which means AI suggestions can be excessive when a system misunderstands writer intention. Novice writers need the confidence to dismiss these, and experienced writers may simply get annoyed. On the other hand, time away from writing can feel like a waste, especially if initiated by tools aimed at inspiration for inspiration’s sake.
Design Space
As an output of my research, I sought to create a resource akin to design guidelines or heuristics that could help designers think more critically about the design of AI systems. Specifically, based on my findings, I created a “design space,” an artifact that demonstrates a key set of criteria a designer should consider when designing a system. In this case, the criteria in this design space was informed by my user research and specifically focuses on how design decisions behind AI systems impact the creative process. The final report included 30 discrete design decisions a designer could make and the implications for each. For the purposes of this project summary, I have summarized the design space here:
Input
Interaction
Basic input systems that rely on selection patterns and only allow a micro level of input can facilitate early-stage, low-expectation ideation.
TextFX is an example of this kind of system.
Systems that require robust, high-volume, description-based inputs may feel more akin to human collaboration but also threaten a sense of ownership in the writer.
Claude is an example of this kind of system.
Output
Singular outputs or multiple outputs that are similar promote convergent thinking. This is especially true in systems that weight results or produce results in high-fidelity formats.
SudoWrite Story Bible is an example of this kind of system.
Plural outputs or multiple outputs that are dissimilar promote divergent thinking. This is especially true in systems that provide neutral results or produce low-fidelity or differently formatted results.
SudoWrite Brainstorm is an example of this kind of system.
A system that is transparent, humanlike, allows explicit calibration and forces the user to integrate results on their own allows deeper collaboration but may appear contrived or inauthentic.
ChatGPT is an example of this kind of system.
A system that offers little transparency or calibration, is transactional in its exchange with the user, or automatically integrates results may be viewed more as a tool than a collaborator.
Jasper is an example of this kind of system.
Reflection
This project is meaningful to me not only because it represents the culmination of my Master’s degree, but also because of what I learned in the process:
I gained a deep knowledge of artificial intelligence, a topic about which I previously only understood at a high level.
In using speculative design proposals in a participatory design workshop setting, I suggested a method for researchers and designers to rethink how they engage non-technical users.
I provided a framework — in the form of a design space — to help designers and product teams consider the implications of AI systems they build.
Something I’m most proud of from this study, though, is my care for and apparent well-executed participant experience. The intersection of AI and creative writing is not just an academic subject or timely debate. It has real implications for creative individuals, some of which are existential and threatening. My study was asking these people to reflect deeply on not only their identity as writers, but also on potential threats to that. That’s why it was crucial for me to consider how to engage participants with empathy, and to choose methods that would inspire positive, constructive critical discussion.