Today, we'll be talking a little bit more on navigating AI, ethics, and user ex user research practice in our office.To start with, we'll be talking about a little bit about me, what happened when AI met when research our research UX research team met AI. We are putting those tools to test, drawing ethical lines, like when did we realize we needed it and finding, like, what works out and where do we go from here. To start with, I used to volunteer as a tour guide in museums across Los Angeles and also was a part of climate designers LA chapter. Currently, I'm also working on a project on human ingenuity versus AI research, examining, like, what are the creative problem solving, constraints in resource constraints environment, which is frugal innovation. We are investigating different, differences in innovation approaches, and the project is called Anotherism. Currently, I lead research across multiple product categories.

Globally, I'll add Electrolux, which is a home appliance manufacturing company. I am a senior design researcher lead for care product line in North America. To start with, the moment AI tool burst out in 2022 onto the research, onto the research research scene, especially when our team started picking picking up. We were both excited and uncertain about, like, hey. How do we use these tools? There was a transition point which deserves a special attention. We all hopped on to, like, you know, learning about these tools, what is it that we can do. We were, pressure testing them. We didn't know. There were no guideline guardrails for us. We were testing out our everyday from writing emails to doing user research. We were testing, what are the possibilities, and we approached this new technology, with our researcher hats firmly on, we initiated with, like, you know, we were initially very amazed by the capabilities.

The speed and the breadth of analysis were stunning, but we also maintained a healthy skepticism. Personally, I was, like, oscillating between being an enthusiast as well as a skeptic throughout this journey. We experimented with various tools, like, you know, general LLMs to start with because they were widely available going from, like, ChargeGPT, Cloud, Copilot, Perplexity, and all of them offer different capabilities and had different limitations for our research purposes. We wanted to pressure test these tools in assessing, like, accuracy, prompts. We wanted to learn about how do we write prompts. On LinkedIn, there were so many, to do list and templates for writing prompts, so we hopped on to that train. We wanted to understand bias and preservation of human insight at the same time.

What became clear was that each tool require different levels of human oversight depending on the specific task. To start with, we used it for qualitative research, and we explored using AI for, like, interview guides and also recruiting screeners. We also tested capabilities for our transcriptions and note taking, which showed immediate practical benefits. Like, you know, before this, we had, we had a note taking person and aggressive note taking for two people currently because of the transcription and the note taking and getting highlights. We, we had a leaner team going forward for each project. For quantitative, we experimented with, like, generating survey questions, analyzing large dataset. The speed of analysis was quite impressive. Like, we could see a lot of analysis happening. And before understanding what is exactly happening on the background, we were trying it out.

Also keeping in mind, generally, as a research team about privacy and data ethics. Like, we wouldn't upload, any identification or private information about our, the people who we spoke to and what the survey was about, but generally on, like, what the data was. Initially, my first impressions were, like, you know, for my job, I think, yeah, I did a great job as a research assistant. I started calling it, like, my intern. Like, since it's since it provided, like, a lot of initial thought starters, it was it wasn't sophisticated enough, to complete to come up with complete solutions without, like, entire context, that we as experienced researchers would bring. But I I think having multiple interns to give us, like, the first draft was was quite impressive. And the acceleration of the capabilities felt like, you know, we were in the movie turbo.

Suddenly, we could move way faster from from what we did previously, and the task that initially took hours to complete were completed in minutes, which was both like, oh my god. This is ex like, how is this possible? Which was exhilarating and disorienting at the same time. As we continued experimenting with these tools, we discovered, like, the importance of we the important limitations and biases and also some of the ethical questions that we had to address. Like, we saw some stereotyping and how some hallucinating at the same time. So one of the critical challenge was, like I mentioned, we couldn't easily determine when these models were hallucinating. They were creating, like, plausible sounding but, like, entirely fabricated information. We didn't know what to do when it lies to us and at where which point in our research that it was giving us wrong information.

This raised, like, serious concerns about our research integrity. It also created non existential journals and researchers and statistics that did sound entirely plausible. So for instance, I was writing a paper and I wanted to learn about a certain reference, and most of the references that it gave me were correct. But suddenly, there was one reference that was entirely made up, and I I didn't know until I actually went down to, like, write the name of the author, the page number, all of it were made of information. And without carefully without if I didn't verify it, and in that moment, I wouldn't understand what's happening, those fabrications or lies or hallucinations could easily would have made it into the research report and recommendations. We observed how AI systems misinterpreted cultural practices, from regions with less representation in the training data. So a lot of the training data we we saw were based in the global North and not in the global South.

So we wanted to compare, like, how those training datas were biased because of the training data, the training it had on and the information. So for instance, a lot of the cultural things in Global South are aren't on isn't on the Internet. It couldn't learn about it. A lot of the text is in the books, and it's it has not made it to the Internet. So there were some biases on misrepresentation of the cultural practices. Also, systems consistently showed bias towards the Western interpretation of global events, which was which was also made us think what's happening here. One of the other challenge that remained particularly troubling, like I mentioned, was the fabrication and lies, which was very, like, you know, compelling and detailed and confident. And especially, like, the amount the level of confidence it has on, on giving us data makes us feel like, oh my god. This is true.

Like, the false confidence was pretty persuasive. The the false confidence was more persuasive than human researchers' appropriate caution. Like, we were a little more cautious, but AI was very, very confident in giving us answers. We noticed that cultural context was also confused by AI systems, like emotional subtleties. Like, suddenly, you say something, in was missed, and it also simplified, like, data points that, were quite nuanced. Despite all these challenges, the positive aspects of our experiments where we saw, like, positive applications were also emerging. It could effectively, like, you know, help us in divergent thinking catalyst.

It was more of, like, in the tech industry, they call it rubber ducking, where they talk to a rubber duck to explain how to solve it. Over here, I think we also used it in the same way. We were thinking with the AI systems on alternative perspectives and which was we're now facilitating back and forth and exploration on variety of topics. So far research preparation and planning also, we AI suggested us methodologies. We selected approaches based on our specific context and situations. So we used it more for suggestion and, like, how do we improve it? How do we push more of the boundaries? We also found it extremely useful for quick summaries, highlights from the interview transcriptions.

Over time, as of now, like, from 2022 to 2025, we have been using AI, which has, like, embedded lead directly into the research softwares that we are using in UX research and also help us in, like, tagging and categorizing, the findings. It was more about, like, you know, suggestions and we we accepting it. Like, they call it having a human in the loop. And the purpose built, of these tools had fewer hallucinations and issues than general purpose AI. That's what we noticed because it was a small language model and also give us recommendations and not give us confidently this is the answer. Drawing ethical lines. One of the things we realized was we needed a structured ethical approach and which was urgently needed because these powerful tools, they were equally, like, you know, robust frameworks for responsible use. We created, like, systematic evaluation of these tools and across different research, context. We reg we rigorously documented each step to ensure transparency and to build understanding, of appropriate applications.

We went from going to learn about the companies, who are these companies, behind these tools, how are their training values emphasized, and also understanding how these systems, were built to anticipate, the limitations. At the same time, we looked into different AI laws and government acts, AI acts in EU. We benchmarked, a lot of legal, and privacy concerns that, all the around the world that were going on. And also we wanted to learn how does AI work. Like, what is machine learning? What is the basics of it going from, like, narrow, AI to general AI? What's like, how do these machines interpret the data from, like, reinforcement to supervised to unsupervised? What does that mean? Also, we came across the black box effect, which was quite, which was a bit of a concern for us.

Like, AI systems often operated without revealing their decision making, logic, and also reasoning processes. So we wanted to learn how did you make this decision, what were like, we see it now happening that, you know, there are references and we can take it and leave it. But back in the day, it wasn't happening, and we are we couldn't understand the reasons of why it was coming up with certain solutions. And with finding out what works, we developed an ethical framework looking into ethics and design. You can also scan this QR code, which was a practical methodology through, like, experimentation after, learning from each of the team, like, you know, what did we do in the UX research process. At the same time, we incorporated concepts of ethics and design, data privacy, and research integrity. We worked to balance, like, human and AI approaches by, asking foundational questions.

What is AI really good at? And also, at the same time, what aspects remain essential for humans in the UX research process. This helped us to identify most appropriate, like, division of labor. What for instance, we were using AI to find out high level patterns in quantitative research. We, and humans, we were looking into, like, what is essential to us having empathy? How are we facilitating? How are we having connections with our users when we are testing a concept? How do we understand nonverbal cues, as well as, like, what what does it mean to be, human in the whole research process, whether it's empathy or, like, you know, understanding subtleties and little expression changes, how do people what is the difference between what do people say and what do people do?

And where do we go from here? We have last five minutes left, and we continue taking steps towards understanding different types of biases that AI has, and we have been developing different bias aware research. For us, we looked into different biases to learn about going from, like, the data bias, like mentioning the selection of, the data that it has over representation of specific groups and under representation of certain minorities. Historical bias going from, reinforcement of past in inequities and also outdated trends that were reflected, looking into social impact, certain reinforcement of social stereotypes when it comes to not just text based, but also image based tools. We had skewed representation of ideologies, which was like the political bias and, also some decisions, were not considered, like, you know, decisions not considering diverse ethical views. It it continuously remains a challenge in our research context.

We did go from, like, you know, using the large language models to understand patterns and, doing the rubber ducking, understanding, and exploring different concepts. But when it comes to our work, going to small language modules and having more controlled AI, in the process. And next, we had was, like, thank you for joining this explorations of AI ethics and research.