AI as a tool in the UX research process

How do designers utilize the advantages of AI without letting it replace them entirely?

Introduction

Attending school in this ever-growing age of AI has given my peers and me a unique perspective on the use and abilities of Language Learning Models (LLMs) as design tools. When I entered college in 2020, ChatGPT was not yet released, and the design industry was just seeing the beginnings of the boom in technology that would completely change the game in the following years.

A few years later, with the release of AI tools like Chat, Gemini, and Claude, the design field began to shift. As a student, I first saw this shift as a threat to my future job security, and I was nervous that I would eventually be replaced. However, as I have come to accept AI as a part of my future, I have started to realize that these tools should not be shunned. Instead, much like Figma and Adobe, they should be adopted as a designer’s tools, rather than being treated as replacements.

Image generated using the prompt: “image of a laptop screen with ChatGPT’s logo, along with Adobe Photoshop’s logo and Figma’s logo on the desktop” by OpenAI, GPT-5.2, 2026 (https://chatgpt.com)

Project Framing:

In a recent class project, I was tasked with using AI to research, ideate, and prototype for a mobile app that seamlessly provides perks and notifications to University of Michigan football season ticket holders. I was apprehensive about starting this project. Maybe, unlike most of my classmates, I was unfamiliar with AI’s capabilities. I had not experimented with it much and was sceptical of its usefulness.

I began the project by writing my own problem statement based on a client brief.

“University of Michigan football season ticket holders need a concise and convenient way to access VIP perks, such as parking notifications, discounts on merchandise, and VIP pregame opportunities, to have a smooth and enjoyable experience on game days.” (Kennedy, 2026)

Then, to test the AI process, I imputed my own problem statement, the client brief, and a short prompt into ChatGPT, asking it to simply rewrite the problem statement. Chat then took what I wrote and expanded upon it.

“Football season ticket holders at the University of Michigan currently lack a unified, convenient way to access exclusive perks — such as advance parking notifications, real-time concessions and merchandise discounts, special pregame field experiences, and social sharing — resulting in missed opportunities and a fragmented game day experience. There is a need for a mobile app that serves as a digital ID to seamlessly deliver these benefits.” (OpenAI, 2026)

I was not surprised by this output. Chat had taken my problem statement and expanded on the existing content by adding more professional and impactful language. However, apart from adding more depth, this process did not save me any time in writing the statement.

Along with my problem statement, I asked ChatGPT to conduct a competitive analysis. This is where things started getting interesting. Chat was able to provide me with an extensive list of potential competitors, and while going through this list, some things surprised me, excited me, and even disappointed me. I found that Chat was able to pull resources from all sorts of sites. Some competitors were very relevant to my project, and a couple of them I had never heard of before. On the other hand, the more I asked for examples, the further off the competitor’s examples became. By continuously asking for information, the LLM became overwhelmed and confused, and began listing irrelevant examples that made the process become oversaturated and confusing.

Interview Facilitation and Analysis

Although I did not see much benefit from the early stages in project framing, I continued with the use of ChatGPT (and UM GPT) for my interview facilitation and analysis. Similar to my first few steps, I prompted ChatGPT using my problem statement, the client, brief, and a short prompt. I asked it to create an interview guide, conducted my interview based on that protocol, and then ran the transcript through UM GPT to be analyzed. Overall, I found that this process of preparing, enacting, and analyzing user interviews has become much more streamlined due to the use of AI. Generating a list of 10–15 questions takes only seconds, thanks to its help. However, when it came to analyzing the data and creating personas, the LLM tended not do as well a job.

As discussed in Jason Godesky’s article, ChatGPT cannot do user research, LLMs cannot understand and translate “user demographics or behavior patterns” (Godesky, 2023). It also builds off of the data that it has been trained on. These two issues combined can often lead to mistranslations and unintended bias.

For example, if a researcher uploads a transcript of an interview or usability test, and asks for insight. The LLM can produce feedback solely based on quotes. However, there is other important information to be gained that the LLM cannot pick up on, such as body language, tone, and past context. Therefore, LLMs cannot be trusted to analyze experiences based on transcripts alone.

On top of that, the LLM may also misinterpret transcripts. I found an example of this when I uploaded my interview transcript. During my interview, the participant mentioned a memory at the stadium: running the Big House 5k. I, the interviewer, know that this event is separate from Umich football events, even though it takes place in the same stadium, and therefore should not be taken into consideration for the project. However, UM GPT did not consider this, and it included that memory as an essential part of one of the personas that I asked it to create.

User Persona created by UM GPT (University of Michigan, 2026):

Text generated using the prompt: “Use this transcript [transcript provided as pdf] to create 3 different user personas that are University of Michigan football season ticket holders. Include the persona’s name, age, job, what they buy at the game, who on the team they support, who they attend the game with, how they get to the game, and any other defining traits.” by OpenAI, GPT-4.1, 2026 (https://umgpt.umich.edu)

This shows me that while ChatGPT is useful for quickly ideating interview questions and organizing the insight, it should not be relied upon for analyzing user data.

Insight Generation

Up until this point in time, the last step in my research process that I have completed is outlining the project scope through a Product Requirements Document (PRD). Similar to my last two steps, ChatGPT was beneficial in saving time and maintaining professional language and organization. When it comes to generating ideas, gen AI can be great for brainstorming. It can be prompted to give extensive lists of suggestions, examples, and potential solutions, and it will come up with as many ideas as needed. This was extremely helpful when organizing my PRD, as there was a lot of changing information for me to keep track of.

However, like the rest of my experiments thus far, I found some downsides to using AI to write PRD. In Prompt Engineering, Boonstra (2024) stresses the importance of being as accurate and direct as possible when prompting AI in order to obtain worthwhile results. By not providing enough context or specific directions, the AI may branch out more and provide suggestions that it thinks are helpful, but in reality are completely out of the realm of the project. This is something that I experienced when drafting the PRD.

Along with this, through my research while doing this experimental project, I have learned that gen AI cannot come up with any ideas of its own. Any suggestions or examples that it provides are acquired from another source or based on its training. This can be misleading at times because no insight that the LLM provides is ever original. Due to that, it is essential that the person who is using the data and suggestions provided by AI goes through them in detail to ensure that all suggestions align with project goals and are not stolen from another source.

Necessity of Human Judgment

These steps of my project and experimentation with AI as a part of the research process have shown me the scope at which AI can produce content at an incredibly fast rate. However, my original question remained to be answered. Will AI take my job?

No. At least as of now, I don’t think that is possible. AI is an incredible tool. It is fast and organized, and it always responds with insight that it considers helpful. It always aims to please. As Godesky states, “ChatGPT is a language model. It never hesitates to tell you that. It can’t analyze data, it can’t consider user behavior, and it can’t come to a conclusion. What it does — really well — is come up with something that a response would sound like” (Godesky, 2023). In the article AI-Powered Tools for UX Research in 2023: Issues and Limitations, Liu and Morgan (2023) expand on this by stating that AI holds bias and misses out on the context of the situation. Because of this, the AI can only build off of what it already believes is true.

Because AI produces responses that it “thinks” the user wants, and because it cannot produce new thoughts, it often produces inaccurate information. This is why it becomes essential for researchers and designers to go back through and fact-check the AI’s outputs. If a researcher lets AI do all the work for them, they are subjecting their product to misinformation, bias, clutter, and poor design choices, which can result in what many people might call “AI slop”.

I found this happening in my own project. In the areas where I let ChatGPT run with the information I provided it, it churned up incorrect and irrelevant information. However, when I tried to rein in the AI’s outputs and edited my prompts, I got back much better results. This has instilled in me the importance of not only checking the AI’s work, but also reevaluating how I used AI as part of the process. I’ve come to realize that AI can be helpful, but it also cannot do everything for me.

Conclusion

Overall, I no longer fear that AI will take over my job. Doing this project has shown me that while AI can significantly speed up the research process through quick brainstorming, idea generation, organization, and writing clarity, it cannot produce its own ideas and cannot understand people well enough to produce products that address their needs. Therefore, AI will always have to rely on real researchers and designers to clean up and fact-check its information. Without real people working behind the scenes, designs will fall flat and become unoriginal. At the end of the day, it is human creativity that makes AI work shine. Good designers should use AI as a tool, and not let it replace them.

Refreneces

Boonstra, Lee (2025) “Prompt Engineering” [White paper] Published February 2024 Accessed: February 27

Godesky, J. (2023, April 12). ChatGPT cannot do user research: Why ChatGPT may produce inaccurate information about people, places, or facts. UX Collective. https://uxdesign.cc/chatgpt-cannot-do-user-research-1c5a34c35abf

Liu, F., & Moran, K. (2023, July 2). AI-Powered tools for UX research in 2023: Issues and limitations. Nielsen Norman Group. https://www.nngroup.com/articles/ai-powered-tools-limitations/

OpenAI. (2026). ChatGPT (February 27 version) [Large language model]. https://chat.openai.com/

University of Michigan GPT. (2026). Response generated by U-M GPT (powered by OpenAI GPT-4.1) [AI conversational assistant]. University of Michigan. https://its.umich.edu/computing/ai/um-gpt