I recently met Preeti Talwai and had a wonderful conversation about AI and how UX research is moving beyond the interface to shape the models themselves. We explored her journey from architecture to Alphabet, why we should treat AI as a raw material rather than a polished teammate, and what a 'brutalist,' transparent approach to AI design might look like.
Without further ado, enjoy!
Kursat: Can you tell us about your journey to where you are today?
Preeti: I actually started out quite far from tech. My education and training are in architecture. I got an undergraduate degree in architecture, and then grad school was more about the history and theory of architecture. I knew quite early on I didn’t want to be an architect. I knew I was much more interested in the social aspects of architecture: understanding people, how they moved through space, and studying that.
I became interested in research very early on in college. I did a little bit of work in cognitive psychology labs on the side, considered a double major, but ended up being more interested in the messy, human, qualitative aspects of lived experience. I assumed that the only way to do what I wanted would be to get a PhD and work in academia.
After undergrad, I discovered UX research while working at Herman Miller for a year. In that work, I conducted UX research for the built environment—workplace research. I was working on next-generation concepts for their office spaces, the Living Office. That introduced me to UX research, futures, and innovation research, which I really loved. So I did go to grad school after that, but with an eye to going into industry. It wasn’t clear to me that that would be the tech industry. I thought I’d work at an architecture firm doing design research.
After graduate school, it was just happenstance that I landed at Alphabet. In grad school, I met someone on the workplace services team at Google. I ended up having a conversation with that person, didn’t end up on that team, but got a message from somebody at Google X who said, “Hey, I ended up with your resume; would you be interested?” It was a contract role, and I took it because I thought it’d be a good experience for a year — why not?
So I went to Google X. Loved it. Working in tech wasn’t what I expected it to be, based on what I knew from my dad working in the industry. So I ended up staying at Alphabet for nine years, from 2016 to 2025, taking a gap year, coming back for a bit, and leaving again.
I sort of fell into AI work, and I never saw myself as a tech person or an AI person. I really loved thinking about the human impacts of emerging technology, which is a lot of what we did at Google X. And I’ve sort of carried that to a couple of different teams: Google X, then the central AI team, which is now Google DeepMind but was called Google AI when I was on it. And then Google Ads, which is quite a different organization and product in many ways, but I continued to do AI work there, too.
Kursat: So you were there before gen AI was a thing.
Preeti: Oh yeah, totally. It was when there were a handful of us. The team I was on in Google AI was the first UX team dedicated to AI in the company's history. I think you spoke to Jess Holbrook, right? We were on the same team, the AI UX team in Google AI. So it was a small group of us who were the first to think about this in an organized way at Google.
Autopilot vs. thinking-by-doing
Kursat: How has AI changed the way you work day to day?
Preeti: Using AI in my work really started post–GenAI. Zooming out from AI for a second, I think about two types of tasks. There are what I call autopilot tasks, where I know exactly what I need to do; I just need to execute the actions, and my brain is sort of turned off. So that’s things like figuring out whether a result is statistically significant, or simple descriptive stats, or cleaning up formatting in a slide.
The other type is thinking-by-doing tasks, which is pretty much everything else. The canonical example is writing. By the end of my time at Google, most of my day-to-day work was writing, thinking, and coming up with ideas. But when I got into the weeds of research, things like identifying themes within the qualitative data set or writing a report were also things where the process of going through them is actually how I figure out what I’m thinking in the first place.
I should mention the autopilot tasks aren’t necessarily always mundane and tactical. Note-taking is a good example of something that’s mundane but not autopilot. It’s a very active activity: I’m synthesizing, organizing, curating, and remembering. So even though it’s a boring task, it’s really a learning-by-doing task.
I try to maximize AI use for anything that seems like autopilot, and really keep the thinking-by-doing tasks human. In doing that, what’s really changed for me is that across all my work, I now break down tasks much more granularly: the nitty-gritty subtasks, to identify which specific aspects are autopilot and I can just have AI do them, and which ones aren’t.
And then there’s also a set of tasks that involve capabilities I didn’t have before, things I’d go to other people for, like design or prototyping. I use AI here pretty discerningly. Let’s say our team needs a high-level concept illustration for a storyboard-based study we’re running, a common use case in my work. It was also common for designers to have no bandwidth, and we might not run the study because we didn’t have those illustrations ready.
I think AI is a fantastic tool when it gives you a capability you wouldn’t otherwise have had. But I also think it’s important to be aware of what is lost when we do that; that self-awareness is important. Because when you work with other people, even to do something as simple as a sketch of a concept, even if it’s purely illustrative, not an actual design mock, just a sketch —
Kursat: Like a whiteboard sketch.
Preeti: Yeah. Or sometimes we’d do three-part things; we’d test storyboards, like, Bob is a user with this problem, then Bob uses this technology, and then this is the solution. And sometimes we just want some illustrations to go with that, to bring it to life. That’s what I’m talking about. On the surface, it seems like that isn’t that big of a deal to do. But when you actually work with other people, you start to clarify that concept, you ask questions about it, have important conversations and realizations, and it can become something bigger.
So even though it’s really empowering for researchers not to have to rely on other functions to put things together, I’m also wary of a world where roles are totally blurred. So I am of the school of thought that it’s important to keep stuff human.
Kursat: When you hand something to AI, what does that look like?
Preeti: The autopilot tasks, as I said, are things like figuring out if a result is statistically significant. At some point in my career, that wouldn’t have been autopilot; it would’ve been important to go through that process myself, but now I just need to crunch numbers, so it’s great if I can put it into an AI tool.
A lot of it is also retrieval. This isn’t necessarily a core research task, but I often know I saw this finding in a report someone wrote last year. I just need to find it and share it with the stakeholder. That sort of connecting the dots became more of my work as I became more senior. And that’s another example where it’s very rote: I know what I’m looking for; I just have to comb through the whole database to find the specific stat. AI is wonderful at retrieving the right thing. I find that quite a magical experience, and it really makes things go faster.
But honestly, for most things, going through the process is really informative for me. So I am very discerning about what I offload to AI.
Opinion dressed as fact
Kursat: Where has AI surprised you these past couple of years?
Preeti: The pace AI is moving at, especially compared to the pace of change in the first eight years I worked on it versus the last couple of years. It’s wildly accelerated. It’s crazy to see what AI struggled with two years ago and now does really well, and to project that into the future.
In my work, I commonly use AI for secondary research on topics. And despite the fact that not much surprises me about the technology (I’ve worked with it for so long that I understand how it’s developed and how LLMs work), I’m always surprised by how confidently it states opinions as facts. Of course, it pulls from publicly available information online, and much of that comes from Reddit and social media. But I’m always surprised by how confidently an AI presents a random opinion from some Reddit user as fact, even when it cites its sources. And it really feels convincing. So I think AI has trouble distinguishing fact from opinion and understanding the reputability of sources. It always reminds me that AI in itself is just another source you have to verify yourself. For all the amazing things it can do, it sort of brings you back to, okay, this has its faults. And that to me is a great reminder.
Kursat: One of our earlier guests, Jess described it to me as an alien intelligence, the ALF character who impresses you and then eats the cat. You still own the fact-checking.
Preeti: Right. And it’s interesting that we often think of AI in terms of the internet or computing. We talk about AI as a tool, and I think that sometimes gives us a false reliance on it. Because when you think of a tool, you think of something that’s going to work more often than not. Versus, I like to think about AI as a material or a source, as opposed to a tool, because I think that’s more honest to what it is. And that resonates with what you’re saying about this alien intelligence: it’s like a thing. Tool feels very cut-and-dry, reliable. If I put ten plus ten into a calculator, I’m not double-checking that; I treat it as fact. But if I think about something as a material or a source or an intelligence, a different category from a tool, then these sort of things become features of the system. This is how the material works; this is how the source operates. And you don’t have that blind reliance on it when you think of it that way.
The collapse: the model is the product now
Kursat: What’s your take on how designing AI products has shifted, before and after gen AI?
Preeti: There’re a lot of shifts. I think the main shift is that AI used to be the technology that powered many things invisibly, and people didn’t realize it was AI. If you think about the advertisements you see online, or the recommendations for what you see on social media, or when you get an Uber or a Lyft, these are all AI. So, researching those things was often the job of a small group within a company. And I spent a lot of my time pre–gen AI trying to justify why we should be doing AI research at all. Research was assumed to be something you do on a thing that has an interface. If there’s nothing for a user to interact with, what are we researching? How do you do research on a model? How do you apply UX research to something like an ad auction? I spent a lot of time explaining and building up the practice of researching the models themselves.
With Gen AI, there’s now very little between the model and the end user. I’m literally talking to the model, seeing fairly raw outputs. And with that, there’s been a collapse: there used to be people who worked on the front-end UI and people who worked on the back end, and now it’s all one thing. And every researcher now needs to be fluent in working with an AI product, because it has enabled users to interact directly with the model and everything that comes with it.
Kursat: Does that change your tools or ways of working? Does the UI kind of disappear?
Preeti: I’ve worked on AI models and human eval for a long time, but it becomes even more obviously important now. Some of the most impactful work of my career has been working with machine learning engineers to shape the inputs and features for their models. And I think that sort of work becomes amplified. It’s really important that the way we train models, and what we consider a “good” model, aligns with what quality, value, and trust mean to real people. That becomes so much more important when the user is directly interacting with the model in such an intimate way, not filtered through various layers.
Before, if I saw a training dataset that was biased in some way, I might help diversify it, review our sources, and improve the human eval questionnaire. But as a consumer, you’ll feel the impact of that work, but you won’t necessarily see it transparently. There was a lot more opacity. But now people talk about the AI; the model is the product. So, shaping the model itself becomes increasingly important.
Shaping the model: golden sets, evals, and “what does good mean?”
Kursat: Can you talk about how you’ve actually shaped a model?
Preeti: Yeah, there are different levels. I’ve done a lot of work defining golden sets, which I’ve found are often defined internally by a small group of people disconnected from users, who decide what “good” means. And that, to me, is often the root of it: really defining what good is, and how we get at that. I believe that should be informed by users, user mental models, and their values, not just what we decide is “good” internally at a company.
And there are other places in that stack. Pre-GenAI, we used human raters (who are still heavily used today alongside other automation). So some of my work was: what questions are we even asking our human raters? How are we writing their survey template? Often, those questions aren’t written by survey methodologists, so there can be many gaps in how raters are asked questions, which then influence the ratings. There are so many parts of this process that UX research can touch, where if you take a close look, you can see where something can be improved to be more rigorous or more user-aligned. So I spent a lot of time working on things outside of traditional UXR to help improve that system.
Kursat: For a design audience, what are the top few mediums for shaping model behavior?
Preeti: Human rater templates, golden sets, evals…they’re all related. Ultimately, the question is: if you’re trying to make a good LLM output, what does “good” mean? And good can mean different things in different contexts for different people. We want to take a user-informed lens on what quality means. What does a relevant answer look like? What does a trustworthy answer look like? So what good is it if it is informed by people — not just people saying, “I prefer response option A or option B” — but by a systematic understanding of people's values and mindsets, and then applying that to ensure the outputs of AI models are both high-quality and responsible.
Kursat: Does that require a different skill set than a traditional researcher?
Preeti: It requires us to be comfortable with a technical understanding, in a way we might not have been before. I don’t think it’s about doing different types of research, necessarily. It could be, but from my experience, it’s more about a different application of what we find, a different language we have to use to communicate those findings.
Often, you’ll hear that UXR should speak the language of product, business impact, or revenue. We often think about influencing product managers, the product roadmap, the strategy.
But you can take the same research and the same findings — for example, here’s what a relevant response means to users — and communicate it to a set of engineers who are developing the models. What’s important is the recognition that they are a super important audience for UXR work, with UXR reaching out into places it typically doesn’t go.
When I was first starting out, a UX researcher had never really spoken to a machine learning engineer or to someone on the human eval team. Now it’s becoming a lot more normalized. So I think it’s a separate communication skill: understanding what’s happening under the hood, being technically fluent and conversant enough to say, “here’s what we found and here’s where it can make a difference from the engineering side.”
Looking around corners: the anticipatory research framework
Kursat: You gave a talk about anticipatory research at the Learners conference. Can you describe the framework and how it applies to AI product innovation?
Preeti: Yes, the framework is really about how UX researchers can look around corners for innovation. We often hear: how do we anticipate organizational needs and conduct research before it’s obvious that research is needed? So the two-by-two: one axis is the thing we want to do research on — is that a current gap in the product, or a future area for the product? The other axis is: does this thing have high executive conviction (leadership believes in it and is bought in) or low executive conviction (leadership isn’t convinced it’s important yet)? And that lends itself to a two-by-two with four quadrants.
I put it together to make it clear that anticipatory research isn’t all futures research. If you work on something that’s an existing gap but people aren’t convinced it needs to be looked at, that’s a blind spot. For a long time in my career, doing research on AI models was the blind spot. I remember going to teams, and they’d look only at the front-end experience (what does this thing look like), with no research ever applied to the model powering it on the back end. So I’d go into teams and say, hey, here’s this blind spot. A lot of my early work on Ads was exactly that: going in and starting up a function that would look at the models for the first time.
But then there’s also work in the future: bets research. That’s: what’s the next big moonshot for this product, for this team? That’s typical futures work.
To your question of how this applies to AI innovation: the framework is really about where UX research is heading in the context of AI product innovation. Like we were talking about, the frontier is advancing so quickly. Looking around corners, anticipating emerging impacts of the technology, keeping a pulse on your users, and identifying pain points or challenges before they become large enough to cause that problem at scale. That is exactly the work of UX research. That’s what we need to be doing. So anticipatory research becomes more critical in the age of AI, because of the pace we’re moving at, the uncertainty of the material we’re working with, and because we’re always on that bleeding edge. We have to figure out what those corners are, look around them, study them, and have a point of view on them. We can’t be complacent. The future is becoming the present very quickly.
Kursat: So the framework helps you stay proactive instead of reactive to whatever the frontier labs announce every two weeks.
Preeti: Yeah, exactly, and we need a language and a toolkit to do it. It’s not enough to just say, “look around corners”, which is what I’ve heard my whole career. I put that talk together because nobody ever taught me how to do it; no one ever had a framework. I hadn’t seen anything that tells me, here’s how you anticipate needs; here are the challenges in doing that research.
But yeah, it’s about being proactive. We often hear, “UX research needs to keep pace with product.” And it always struck me as weird: why are we constantly running this race to keep up? Which feels like a very reactive mindset. Versus, if you free yourself from that and say, okay, I’m going to look a couple steps ahead, because the product is going to get there, but I need to be there first and understand it, so that when the product comes to that point, the research is ready. That way you break the cycle of running after your stakeholders and trying to do things faster and faster. So we’re being quick and proactive, but not racing to catch up.
Kursat: When you drop into a team, do you start with the two-by-two?
Preeti: It’s more about identifying which of those quadrants the work you want to do is in. I think researchers have a good sense of what we should be doing; many of us have a good pulse on the product, sometimes better than anyone else, and know which questions and work to pursue. Where we get held back as a discipline is feeling pressured to focus on other questions, because we’re told to by leadership or stakeholders.
So the quadrants are about identifying what’s going to be impactful and giving researchers a language to name it. In that talk, I gave a couple of questions to think about. For example, what is one set of stakeholders on your team who’s never worked with UX research? Maybe it’s some engineers. That’s probably a blind spot, an area where you could be using UX research, but you’re not. Or, what’s a theme that keeps coming up in meetings with leadership, but no one has a great answer for and no one’s really pressing on it? That could be a really good topic for some bets or frontiers research.
As UXRs, we have this running catalog in our heads of those things, but we don’t always have the vocabulary to express them or a method to go after them. So I don’t literally make that two-by-two or share it with stakeholders; it’s more a way for me to organize my thoughts and be like, here’s a place where I really think we could apply UX research, and then understand what it’s going to take to pitch this idea, get buy-in, validate it, and scale it into a research program. It’s more of a researcher tool to get the confidence to do the stuff we know we should be doing.
Kursat: It’s a mindset plus a language, even the permission.
Preeti: Yeah.
The teammate question — and accountability
Kursat: In my work, I’ve been exploring the AI-as-teammate metaphor, borrowing from team literature to treat human–AI teaming as a design challenge. What’s your point of view on tool vs. teammate?
Preeti: It’s interesting because I actually think about AI as a material, and maybe this is my architecture training. There’s a famous architect, Louis Kahn, credited with saying, " What does a brick want to be? This idea that the material wants to be a thing, wants to act a certain way. You don’t force it to bend to your will. It’s something you build with; the material is the character. You’ll hear people talk about movies this way, too: for movies shot in New York City, the cast often says New York City is a character in the film. That’s how I think about AI.
I struggle a little bit with “AI is a teammate” – although I agree with the sentiment, which is maybe the same sentiment I have when I think about materiality. It’s not just a calculator; it’s contributing, generating, creating, and adding something to the system. Where I have questions about that metaphor is regarding accountability.
Kursat: Yeah.
Preeti: I often think about errors or novel contributions by AI, and what happens to accountability when we think of it as a teammate, a third person in the room? I don’t have a clean answer for that. It’s just something I wonder about with that metaphor, which is why I tend not to think about it that way in my own work. I don’t know if you have thoughts on accountability, but that’s the thorniest part for me about having AI as a teammate.
Kursat: Accountability comes up in almost all of my conversations too. I keep landing on something between a tool and a teammate: a new kind of collaborator, without the full teammate role. Building on your material framing: what material qualities does AI have that are different from other materials?
Preeti: To me, and this connects, you mentioned industrial design, so I thought of Nick Foster, an old colleague of mine at Google X. He’s written about this concept; he may use the word material or something else, but it’s the same idea. I agree with a lot of what he’s written. For example, right now, when AI hallucinates, we’re like, oh, that’s bad, that’s wrong, it was inaccurate, it betrayed my trust, I can’t trust it. But if you think about it as a material: if I’m building a brick wall, I can’t be upset that brick is hard and inflexible. You cannot expect it to behave like something else. Or you can’t be upset if something you’re building out of a porous material leaks. Or if you think about fashion or clothes design: you work with the material, with an understanding and a respect for the material, rather than trying to make it something it’s not and then being upset about it.
Another characteristic is that a single input won’t produce the same output; that’s just not how it works. Also, with machines that can self-improve and learn, how do you work with something that learns over time, takes feedback, and can be trained on it? That — the hallucinations, all these things that are quirks of the system — how do we research them, work with them? It’s not to say we should accept inaccurate outputs and run with them. But accepting that this is how it’s going to be when you work with AI, to me, is a really important part of it. Versus treating it like a calculator where you’ve got to get the right answer every time. I just don’t think that’s the right model; our expectations won’t be met, and then we’ll have a false reliance on something we could address more productively in other ways.
Kursat: It reminds me of wood: the material behaves differently over time than when you first bought it, and you accept and work with the constraint.
Preeti: Yeah. And you can also think about (this is something I’m starting to write up and think more about myself) what would a brutalist AI look like? That’s another aspect of materiality I think about a lot. I’m fascinated by brutalism in an architectural context. But what does it look like for a material to reveal itself? Exposed beams, not putting a facade on it, not painting it. What does that look like for a model? We do have part of that; you can kind of see what the system is thinking. But a lot of it is couched in very anthropomorphic language. The chatbots often say things like, pondering it/ Sometimes, when something goes wrong, the AI says, “Oh yeah, I messed up; I was just eyeballing it.” And I’m like, you don’t have eyeballs; what does this even mean? What would it look like if we had a system that truly revealed itself and trusted users to handle it, rather than putting a facade on it or dressing it up? Or if you think about reinforcement as a metaphor, like reinforced concrete, what does it mean to have AI as a teammate or a material, but we need to reinforce it in some way? These metaphors are something I’m starting to think about and write about.
Kursat: Are you thinking at the interface level, at the behavior level, or both?
Preeti: It’s not an idea I’ve fully articulated yet, but for example (and this is quite hard to do technically) I think about revealing some sort of confidence score for what the AI is saying — like a version of the log probability or consistency score. Right now, AI just tells you something as a fact, and always confidently, even though on the back end there are various ways of assessing its confidence; we don’t reveal that to the user.
There’s a lot of talk about how we can get users to trust AI. And sometimes I feel like that’s the wrong question. Building for trust, designing for trust: it’s a noble intention, but sometimes it feels like we’re engineering trust. And I think about brutalism as the opposite: you have to trust that when you show users something maybe complicated, unexpected, not anthropomorphized, just really raw, they can handle it. Brutalism, to me, it’s about honesty. That’s how it started, material honesty. And I think about, instead of designing a trustworthy AI, what does it mean to design an honest AI?
Kursat: The confidence piece comes up constantly in my projects.
Preeti: Yes, I just wrote a piece about this… log probabilities and consistency scores can be translated into “this is medium, low, high confidence.” Or, “here’s an answer, but this should maybe be human-reviewed.” Things like that can be helpful. It’s technically difficult, and it’s not straightforward. But right now it presents everything as if it’s 100% confident until you call it out, and then the AI is like, “actually I messed up.” That pattern should not be happening.
Kursat: I guess it’s not too hard to do; maybe the product just doesn’t want to appear weak or unreliable. But I like the thinking; hopefully, more companies will have the confidence to show the AI’s weaknesses and vulnerabilities.
What she changed her mind about: living in two time zones
Kursat: What’s one belief about AI and design research you’ve changed your mind about?
Preeti: This comes from my personal experience. I worked on AI for many years with this underlying assumption that when you’re designing for a person whose life is going to enter the future all at the same... that different parts of their life are going to be modernized at an equal rate. That’s the assumption I had. And obviously, as designers and researchers, we talk a lot about adoption curves on the population level. It’s well understood that emerging technology, including AI, does not arrive everywhere equally; the future is unevenly distributed, we know that. But on an individual level, we often tell these user stories where it’s like, oh, this person’s life is just going to exist in the future.
Last year I had a really major surgery. I lost my entire large intestine. So I now live with an ostomy and an ostomy bag. It was crazy to me that I’d wake in the morning and snap on what is basically an ostomy bag, a glorified Ziploc bag; it’s odor-free and whatever, but it’s a plastic bag I wear on my body. And it feels so primitive, like the bare minimum needed for dignity if you have a condition like mine, and then I go to work, and I’m working on sci-fi stuff.
And it just made me realize that the more AI advances, the further apart these different poles of our life will shift. There are some aspects of our lives that will feel incredibly primitive; this is one example, but probably everyone’s life has something like that, something that feels stuck really in the past. And then there are some parts, like the things that AI can handle, that feel super futuristic. And we kind of live this anachronistic life. At least, I feel like I’m kind of living in two different time zones technologically. And that’s kind of jarring.
And I’ve never considered that reality when designing or working on AI products. I think we need to be more sensitive to that reality, that fragmentation, when we’re designing. And I don’t have an answer to what that means, but it’s a consideration I think about now that I never thought about before.
Kursat: It comes down to lived reality: we’re not experiencing it evenly.
Preeti: Yeah, it’s on a very individual level. I’m very used to thinking about next-billion-users and adoption curves; on a population level, it’s obvious. But on an individual level, it’s just not something I’d considered or felt so viscerally.
Advice for people navigating the AI wave
Kursat: What advice would you give designers and researchers navigating this: junior folks, and mid-career people trying to upskill?
Preeti: There are so many aspects of navigating this. There are the job-security concerns. There’s the reactive feeling of having to catch up and upskill. I think the advice I’d give is probably the same I’d give anybody in a moment of transformation or flux, whether or not it has anything to do with AI. And that is: being really clear on your values, motivations, and purpose for what you do. Instead of being attached to a certain job description or a company or a skill or a technology, really figuring out your “why”. It seems a little cliché, but figuring out your why and not compromising that — really holding that line. Knowing what you stand for, and why you do what you do.
The focus is often on AI: how do I learn the skill, and what will AI do to my job? And I think we ought to do a lot more introspection. What am I learning about myself? What are my feelings about AI teaching me about myself and my values? We will need to be flexible in how we define ourselves and our identity. But part of doing that is being really clear on your values and your purpose. And researchers especially, we tie a lot of our value and our worth to that title. If we dissociate from that a little, and instead think about what values and qualities are really important to maintain, and hold the line on those, then it becomes clear: where do I invite AI in, what do I do with it, what’s my metaphor for how I think about it? It starts with understanding ourselves better first.
Kursat: That’s a beautiful way to conclude our conversation. Thank you so much Preeti, this was such a rich conversation.
Preeti: Thank you for having me, this was fun.
And this was a wrap for this issue. Until next time, take good care of yourself, and your loved ones.