This short paper explores the use of generative AI for conducting user research in interaction design, more specifically it will examen the rehabilitation experiences of elderly patients at a Danish physiotherapy clinic. Three personas representing potential users were generated in order to perform simulated interviews. A total of twelve interviews were conducted: the first three using synthetic data and the remaining nine grounded in actual user research. It may be concluded that, data-driven personas are more effective in capturing nuanced user needs, however reliance on AI-generated personas without empirical data introduces risks of oversimplification and misguided design decisions. Thus, generative AI ought to be integrated as a complementary tool, used to assist the designer in their process alongside expert validation. User studies, are crucial for understanding individual motivations, challenges, and behaviors. Designers conduct user research in order to capture tacit knowledge, that are experiences which are difficult to explain or communicate. Personas can be create to synthesize the data from interviews and depict a generalized version of the users. Moreover, interviewing users can be resource-intensive and sometimes inaccessible due to geographical or contextual constraints. Generative AI presents a novel tool for simulating user studies by creating dynamic personas and facilitating qualitative analysis. In this study, AI-generated personas based on anonymized data from Fysiofresh a Danish physiotherapy clinic, will be used to explore generative AI in the design process.
Initially, three generic personas were generated by ChatGPT based on a data structure suggested by Nielsen and Hansen (2014). However, after consulting with Fysiofresh, anonymized real-life data was provided to create more accurate personas, thereby potentially mitigating biases. This data driven approach allowed the creation of three personas that represented elderly Danish individuals considering Fysiofresh’s rehabilitation platform. Nine interviews with the data-driven personas were conducted over three days. On the first day, participants answered general questions about their treatment experiences and needs. The second day featured UI co-design sessions, where personas reviewed wireframes and proposed additional features. The third day was dedicated to gathering critical feedback, addressing optimism bias in line with Jeppsson’s (2024) recommendations.
Following Graneheim and Lundman’s (2004) approach to qualitative content analysis, two additional "expert personas"—a physiotherapist and a UX researcher—were used to analyze the transcripts. This was conducted with the OpenAI o1 model, leveraging its reasoning capabilities for categorization and thematic abstraction. A word cloud, generated using ChatGPT-4o, was subsequently used to visualize key themes. The three initial personas produced by ChatGPT-4o were notably lacking in personality and tended to give stereotypical answers. Relying solely on these “generic” personas could thus enforce bias and risk the development of features poorly aligned with actual user needs. Thus, while creating personas without empirical data may save time and allow one to test the capabilities of AI-driven tools, such an approach often results in verbose, overly optimistic characterizations that lack genuine personality and nuance. Personas grounded in real user data however, seemed to yield more nuanced, contextually relevant answers. Furthermore, these personas offered more developed characterizations, most likely due to that they are built directly upon user research, thus providing suggestions for improvement that align with their personality.
When reviewing early wireframes, the personas highlighted several accessibility concerns, including small text, ambiguous icons, and low contrast. However, the text size was already set to 20px, exceeding accessibility standards (Accessible Web, n.d.). By and large the AI-driven persona proved to be a valuable tool for brainstorming and early ideation, but can be prone to overly optimistic conclusions. Herein the personas where prompted to adopt a negative perspective on the wireframes, which proved especially insightful as it yielded constructive feedback that was, in many instances, accurate and actionable. The personas feedback drew attention to specific interface issues, including problematic color choices. Ultimately, being challenged by negative personas offered a valuable opportunity to refine the design in ways that more closely align with authentic user needs and expectations. The analysis by the two “expert personas” (a physiotherapist and UX researcher) identified five key themes in the design of rehabilitation tools. Personalized, empathetic care was prioritized, emphasizing tailored interactions. Family involvement was also key, improving motivation and treatment adherence. Accessibility and usability were important, with a focus on clear instructions, larger text, and simple navigation. Transparent communication, including clear treatment plans and cost details, built trust. Finally, emotional engagement, through positive reinforcement and motivating feedback, was crucial for sustained participation in rehabilitation.
This study demonstrates that data-driven personas yield better results opposed to those generated solely by ChatGPT. Adjusting prompts to elicit negative responses proved beneficial, reducing optimistic responses which could further elicit biases as it agrees with the designers statement, opposed to challenging their assumptions. Although generative AI systems like ChatGPT show promise, they cannot [yet] replace user research on actually humans. Furthermore, generating additional personas could assist the designer in their analysis and may hold potential, however further research and refinement are necessary. Generative AI can elevate the persona creation process, making personas more dynamic and data-driven. This approach can reduce initial research costs and facilitate early ideation (Nielsen & Hansen, 2014). However, biases in training data, sensitivity to prompt formulations, and the risk of hallucinations must be taken into considerations. To mitigate these downside I recommend either reaching out to experts in the domain of research, to validate or performing triangulation with real user research. Notably, capturing tacit knowledge extends beyond merely gathering answers to questions (Sanders & Stappers, 2008). As this demands eye contact and body language in my own opinion.