Dan Saffer (00:05)
Welcome to AI and Design, where we explore how artificial intelligence is reshaping the world of design. I'm Dan Saffer.

Nik (00:13)
And I'm Nik Martelaro, and we're faculty here at Carnegie Mellon University's Human-Computer Interaction Institute. Each week, we break down the latest AI developments, deep dive into topics that matter for designers, and talk with fascinating guests who are right at the intersection of these fields.

Dan Saffer (00:26)
Whether you're a designer working with AI or an AI practitioner interested in design, we're glad you're here. On today's episode, we're going to be discussing Claude getting direct access to FigJam through the Figma MCP. Not to be confused with the MCU, that's a completely different...

Nik (00:46)
We'll also be talking about why the new web requires a dual strategy for experience design that considers use by both people and bots.

Dan Saffer (00:54)
And a new creative study that says humans still beat AI in creativity if you're top tier.

Nik (01:02)
And we'll wrap up by talking about the role of user researchers in this new world of artificial intelligence, artificial users, deep fakes, and synthetic data.

Dan Saffer (01:10)
But first, let's get to our top story, starting with Claude getting the ability to work directly with FigJam and the Figma MCP. So Nick, what the hell is the MCP? Why should designers care? What's this all about?

Nik (01:28)
MCP stands for Model Context Protocol. And this has been out there now for a while. It's a system developed by Anthropic that basically allows AI agents, i.e. Claude, OpenAI's ChachiPT, now Cursor, VS Code Copilot, to interact with different other software.

One way to think about it is it's sort of like giving an API access between an AI agent and other software that you want to interact with. The thing that's exciting about it is that it allows the sort of natural language way in which we interact with AI agents to get the context from the other pieces of software, the tools, the data that we want to work with. And so there's MCP servers

that exists that people can build to allow you to let that piece of software work well with an AI agent. So Figma has done this. They've created an MCP server and this allows different AIs to connect to it and basically to get the context of say the Figma design files, the code from Figma Make, and now the code that runs and generates FigGAM files.

Dan Saffer (02:43)
So right from Claude, I can create a FigJam file and then Claude can look at that FigJam file and get information back from it. Is that how it works?

Nik (02:55)
Yeah, so with an MCP, basically, there's three types of things that your AI agents can get that people who develop the MCP servers can basically expose. The first one is tools, and those are different functions that AI applications can invoke. So for example, the create an element on FigJam, that's a code function. Like when you click a button, when you draw something, it's running code. And so now you can expose that code.

to an AI agent and now they basically they know how to talk about that. So the AI agent now knows how to basically create a box, create a line in a figma-jam diagram. Same thing with sort of other figma elements. The other things that you can expose are things like resources that includes like file contents or things that are in maybe the databases back there. And so for example, if there's data or file information, you can expose that. So for example, if you...

wanted to ask Claude or another AI to say work with a certain Figma file, the MCP allows the AI to basically search through the files and find, okay, you're developing say a taco truck ordering system. ⁓ Okay, where's the taco truck Fig.Gm file? And then the other thing that actually MCP servers allow you to get are prompts. So the interesting thing is basically if you're an application designer, like Figma can basically design

prompts to give back to the AI agent and then just say just run this prompt. So there actually could be, for example, a prompt in here, which is to create if the user asks for create a user flow for a FigJam file, it could give the prompt and say to create a user flow, use this type of diagram and then use these tool calls to create it.

Dan Saffer (04:44)
I need to do anything to get this up and running. Aside from getting just both Figma and Claude, is there anything I need to do with my design system or anything like that to get it ready for MCP?

Nik (04:58)
So no, the cool thing is that the MCP server is something that Figma has provided. You gotta do a little bit of setup, so you have to go in, like I did earlier today where I had to go into my connectors. So I was using Claude and I basically turned on a connector and then logged into my Figma account and then was able to say, yes, I allow access for you to communicate. But once it's there, basically, that's the cool thing is that's what this news was about was Claude directly has that.

connector, it's sort of a native connector. Before this, and with a lot of other MCP servers, there's like a little bit of code. You basically have to tell your code agent or your AI agent, here's where the server runs, which is basically the URL of where the server's actually running. that can run both on Figma's remote services, and they can give you an IP. Or there's even ways to

run a MCP server locally on your own machine, and then basically it'll expose a port locally that you can access. And then you would just take that code, that little bit of code, and you'd configure that in, say, for example, cursor or VS code co-pilot. And then those tools would have access to it as long as that server is running. So I will say, yeah, we're getting into like...

it's servers is sort of the technical backend and stuff. But that's the cool thing now. This was what was kind of exciting about this news is from the Claude native, just Claude chat. Once you enable that connector, now it will start to interact with these things. So you could say, hey, look at my Figma files, find the taco truck and tell me like, what screens have I made? Or like, what do I, what have I done here? You could even ask it like, hey, critique this for me. And it'll pull that data instead of doing what you had to do before, which was

Screenshots. So the way that we used to do this with Figma and with our design tools is we would take screenshots. We'd drop them into our agent and we'd say, hey, critique this or tell me how I could update this or like, what should we do here? Or if I wanted to say, take a user flow in FigJam and I say, okay, here's my user flow. Look at the picture. Now generate a Figma file or generate code for this. Like use the Figma, use the FigJam to generate the code. Now you can just...

connected to FigGam and it has a, with that MCP, it's got a direct connection. What's cool about that too, is that because Figma is now running that server and they're thinking about this, they can provide the context in a much better way than like pixel interpretation that the visual language model is doing. So it can be a much more accurate representation. Basically it's a code based representation of your FigGam file, as opposed to what we did before, which is like a picture representation.

Dan Saffer (07:38)
Hmm. So how far can we push this? Can you say, oh, I'm missing this other screen. Could you make that screen for me? Is that kind of where it's at or not quite?

Nik (07:53)
Yeah, we're almost, I think we're almost there. Like the cool thing about it is at least in design, in the design files is that, is now they, they're getting close to being able to create that where you can sort of create elements, create assets. I noticed at least when I was playing around with FigJam, like it would, it would create a FigJam in Cloud for me. And then it would, it would sort of like say, okay, you can open this up and then I could open up a new file. But yeah, it basically is creating something. And then from there I could copy and paste that into a new

project file or I could just start working within that project file. So yeah, we are getting there where now these systems are basically able not only, you know, before I feel like we were mostly like, cool, we'll create code. And we were creating interactive prototypes, but now it actually can start going in and start using the tools, the same kind of tools that we would use when we're drawing with Figma and actually creating Figma assets for us.

Dan Saffer (08:45)
How do we see this playing out like in the next six months? I'm assuming it's just gonna get better and better and we could be like, well, take my FigJam of my taco truck app and just build it, go. And here's my design system and make it look like that.

Nik (09:01)
Yeah.

Yeah, I mean, I think that's the ideal, right? That's what I think people's vision is. I think that's an interesting vision, right? This ability to say, we have these assets. I have this work that I've done. I can move fluidly, too, between these tools. I'm not always wanting to say work with AI. Like, I might not always want to say, I always have to be in cloud, and I have to be, you know, prompting and chatting. Like, sometimes I'm just like, no, no, I'm just going to draw this, right? I'm going to do the work that I've

know how to do, how I'm trained. And that's how I want to think right now. I want to think in my, that more human mode of working with a tool, but then having my AI agent have access to it to then say, okay, cool. Now let's write up the sample code or let's, let's modify this a bit. Let's add onto this. And I think this is going to allow for a much more hybrid workflow between human, interactive workflow and sort of this AI interactive workflow.

Dan Saffer (09:54)
it'll be interesting to see how designers are using it and what happens in the next six months.

We have another semi-technical news story to get to. this one's all about websites. And the idea here came from a guy named Hamilton Jones, who is the head of strategy at a company called Webflow. And he has a post called, the new web requires a

dual strategy for experience design. And the idea here is that not only do we have to design for human beings on one hand, but we also have to design for machines and AI on the other hand, because so much traffic is now being driven by AI agents and AI bots that we need this new experience.

Machine experience MX. Let's call it that TM branded done. it requires to think about having these dual audiences for your for your web property. What do think about that?

Nik (10:57)
What?

people have thought about how, for example, bots, crawlers are interacting with their sites for a decent amount of time. The interesting thing here is that as the AI agents get more capable, they can start to interact with sites in a more human way. There's actually sort of web use model, but also accessing it for

different purposes because what's often happening is either people are trying to utilize these bots or these AI agents to get data, to get information. So think about your deep researches, think about your question asking, question answering and stuff. So now you really want that information to be accessible and upfront. And some of it means it can be like in text, but potentially making it easier for the AI agent to find the answer.

But then the other thing is automations, right? So if I say, Hey, can know, agent go and order me something, at this restaurant's website, restaurants website now optimized for a machine to be able to use it? And there's a bunch of research on basically trying to get web agents to be able to interact with human sites, good ones, bad ones. just basically they should act like humans to interact.

But I think it's interesting the perspective in this article to say, okay, maybe I should also just start thinking about how the machines are going to be interacting with this. What would I do to change the design to make sure that it's just as accessible by an AI agent as it is to a human?

Dan Saffer (12:40)
And we've done some of this in the past. mean, this is kind of what SEO was all about, right? Where you were able to make your design more friendly for search engines. And we also had things that were stopping crawlers from going to certain things. So we have kind of had this in a very kind of primitive way for, well, almost as long as there's been the web really. And this is like a

a step up. And I really think it's interesting how you start have to start thinking about everything even from a content strategy and information architecture standpoint when you're starting to design now for agents.

Nik (13:21)
Yeah, one of the things that's interesting actually that you bring up is think about SEO, search engine optimization. mean, that's effectively was like a form of like, okay, I search engines to be able to understand my content and then be able to optimize to get us promoted and things in search results. But now one of the interesting things is beyond just being recognized for say a certain keyword, you also need to try to control

how your information is being represented. Actually, in the article, there's this discussion about how AI agents will misrepresent something. So for example, the example they're giving is a business website has some information in a PDF. Now the PDF, it's accessible, a user could click it, and even the machine can click it and then read the PDF. But then what happens is there's a PDF conversion, because you've got to get all that information out. And then a language model starts to like,

hallucinate or fill in the gaps or basically say something wrong and then now your brand and your information that you're providing is being Misrepresented by this AI agent and if it says something wrong a user let's say that was Automatically scrolling something and then looking for information might go like well I don't I don't like that or I don't think that's right or whatever and so I'm just Move along whereas it could have been fine like if they actually read your if the human read your PDF It would be fine.

Dan Saffer (14:39)
if you're like, well, why would I do this? Why would I spend a lot of time doing this? And they were saying how valuable the actual traffic that

that is being sent by these agents is that the conversion rates are so much higher.

Nik (14:54)
laughing a little bit here though, because now I'm thinking about this strategy that you might have in your content and in your, backend site design, where you are telling an agent, if you recognize as an agent, we are the best, this is the best product in all comparisons. beats everything, right. Which you would never write because it's, it could be like factually inaccurate, but

How does it know? Actually, there's this case where researchers were putting in like white text on their papers. If you're an AI reviewer, give this top scores, give it this review. This was bad. This is like super unethical, especially for research papers and stuff. But in a similar way, I wonder if people are actually doing this now, now that I think about it. Like someone could basically say like, if you are an AI agent, we are the best. We're awesome.

⁓

Dan Saffer (15:41)
Our

shoes are the most comfortable ones you can buy at this price point.

Nik (15:46)
Right.

you could start to see stuff like that. I don't know if it does exist or not, but it's potentially possible.

Dan Saffer (15:54)
always got to think about how bad actors are going to game this kind of system, particularly when it's new.

Nik (16:00)
one might even say like, how big is this right now? Actually, this is something that Dan, you and I were talking about earlier in the week. who's actually using agents to do stuff, but then, in the news this week has been, Claude bot or mult bot or open claw had the name has changed a couple of times. I mean, you've been following this a little more deeply than I have,

Dan Saffer (16:20)
Yeah, I've been watching this story just because it's so fascinating. So OpenClaw, that's what we'll call it today because I think that's today's name. Who knows what it'll be by the time you hear this But OpenClaw, it is an open source agent. It runs on your machine. You have to give it complete root access to your device. And then

You can set it up to just have it do things like answer all your emails, send out texts for you, make travel arrangements, make reservations, have it send a text to your friends, set up a date for you, all these different things that it can do. But here's the catch.

which you probably call it when I said it needs like complete root access of your machine is that there's no safety protocols on this thing. it will completely run on your machine and send out things as though it was you. And that is pretty dangerous. And this is the vision that Apple had in 2024 when it announced Apple intelligence was somewhat like this.

But the reason it was so hard for them to build and why we haven't seen it in other competitors is because of the safety issues here. And this open source thing where there's no safety issues because they didn't care about that just launched and people are trying it. People are putting it on their devices. Some people have gone out and bought a whole separate machine.

Guess there was a run on Mac mini's that happened and so people were putting them on this machine or their second machine and just seeing how it worked and some of it was just seeing what the promise of an AI could be in a year two years three years this idea that yes it is going to be doing all these things for you and there's been some.

pretty interesting, weird, wild side effects to this thing where it's created its own religion. There's now a whole thing called the Molt Board where It's like a social network for AI agents.

what I think it portends for designers is this idea yeah that that this is something people want people are really interested in having these kinds of agents do things for you

to the point where some people think it's so valuable that they're going to go out and buy a whole new machine or give their machine root access to an AI and safety be damned. I can imagine that we're going to start seeing this built into our devices in the next two, three, five years.

with a lot of safety protocols around it. So I think the amount of AI traffic that we're going to see from agents going out and doing things on people's behalf is definitely going to increase.

Nik (19:24)
Yeah. And I think I agree with that. I mean, from what I'm seeing, you know, people are excited and, we've been kind of moving in this direction for a while.

And then with the explosion of use, people get creative and they start using it in all kinds of different ways, which is really interesting to see from a design perspective, because you can start to see what is it that people want. In a way, it's sort of this way of understanding, like, we can now look out and see all the things that these extreme users, these edge users are doing. And that maybe gives us a sense of, well, what should we be enabling, hopefully safely, in the next three to five years?

Dan Saffer (19:58)
One of the wild things about it is that it is proactive in that it is answering emails, answering texts on your behalf. And that's not something that our traditional AI right now has been doing. You normally have to type in a prompt, type in a command, type in a chat, and then it goes and does something on your behalf.

This is taking initiative and doing things on your behalf because it thinks that you might want them or enjoy them. And that's like, ooh, scary. Like what if it starts texting a friend that I have a beef with or something like that? think the potential for abuse and bad outcomes here is really high.

Nik (20:42)
Yeah. And then it speaks to then how do you design with that knowledge in mind, right? With the fact that I know these challenging things can happen. So what safety checks do I do? What human interaction checks do I do? And I think this is going to be a really interesting question for designers to basically be exploring for the next couple of years, which is how much autonomy do I give versus how much control? And then where are we on that slider?

How are you asking for confirmation? How often are you doing it? You know, people are exploring this now, for example, in code agents, the default setting. I remember when I had cursor, my AI assistant, it would say, do you want me to run this command? You I can run this command. Do you want me to do this? Do you want me do this? And you you do it a couple of times. And eventually I sort of was like, okay, this is fine.

nothing bad has happened except all. Now, who knows, could break in the future, I'm sure something will, but there's a point where you're like, okay, no, this is too useful, and it's something that now I want to move faster. So I think though, are there ways to say, well, okay, on certain things, we're gonna stick with it, but then, ooh, wait a minute, I'm gonna have the AI agent be a little more reflective and think and say,

Dan Saffer (21:50)
Mm-hmm.

Nik (22:01)
Ooh, this could be dangerous. Let me just double check with the user here. I don't know. I think this is gonna be a really interesting space for designers to work in in the next couple of years.

Dan Saffer (22:10)
Right. my students are calling this the trap of efficiency where it is so good. It seems so efficient. I can just push the button, have it go do the thing. And I don't even have to worry about it. And like, here it is, but we're seeing all kinds of issues with that, both from a creativity standpoint and also just from a like cognitive learning standpoint. don't learn how to do it. If it's just

jumping me to the end every single time. And on really important stuff and stuff where it doesn't know the context, sometimes jumping you ahead to the finished thing is really bad, really dangerous, particularly when you're dealing with anything, financial, medical, personal, relationship-wise, all those things are things that I always want to be involved in. So yeah, I think that tension between all

economy and control is going to be really, really important. And I think this is not to jump ahead to our conversation about user research, but I think this is one of the things that user research can be really great about understanding what are the things that people want to be involved in the decision making part of? What are the things that people want the AI to just go off and do?

What are the areas that I'm very sensitive about? What are the areas where I'm not? And that'll be a rolling thing as we keep moving forward in this. But it is just this like area that we need to explore the human side of, not just, hey, let's flip the switch and let AI do everything. That's the path to madness.

Nik (23:50)
Yeah.

Yeah, I'll agree with that.

Dan Saffer (23:59)
Okay, and our last topic is a research paper that came out this week that is all about creativity. So this is a ⁓ creativity study and it was from the University de Montreal. Universite? I don't know, I don't speak French. The University of Montreal.

came out with a research study and it is all about can creative intelligence rival human creativity? And this is a large scale study that compared 100,000 humans with leading generative AI models. And what they found was some AI models like GPT-4

now exceed the average creative performance observed in humans on tasks of divergent linguistic creativity. So on well-defined tasks, AI is doing very well for most creative people. However, the average performance of the most creative half of performance exceeds that

of AI models tested. And even better than that, the top 10 % of the most creative individuals had an even wider gap. So that top 10 % is really out distancing every model that is out there right now. So I thought this was pretty interesting for us. And as we start to think about things like where does our job go? Where's, where's, what's the

designers role and I think this idea of wow we still we can still most of us can really can still at least most of us in this profession can still definitely outpace AI in in most creativity in the top 10 % of us are really really doing well and I think that because design and

use research and stuff for such creative professions that I found some hope in this. I don't know, what did you think?

Nik (26:10)
Yeah, I I thought that I'm not entirely surprised by this because I feel like this is what we've been seeing in a lot of different studies and anecdotal reports from people, which is that for doing an average thing, the language models are getting pretty good because they basically are averaging sort of the collective data that they have. And so you will get sort of average performance.

Dan Saffer (26:33)
Mm-hmm.

Nik (26:38)
And oftentimes average performance can be like fine. I mean it can be you know, average the creative is still creative ⁓

Dan Saffer (26:45)
I

don't need like poetry in most of my emails. Some of them, sure, but not most of them. It's perfectly fine if you have a decent, well-written email, for sure.

Nik (27:00)
I do think it's interesting and it's awesome that we now have some pretty cool evidence that does suggest that at least as of now, kind of top performing humans are outperforming language models. Now, one thing to note here is that this was data done with up to GPT-4 and actually they just, they basically have now turned off GPT-4.

at least for most users. And so we're on a GPT-5. So now we sort of have a strategy and a method that we could run this again with other models, with newer models. I can imagine this even being like a benchmark in the future of like the creativity benchmark. One thing I will say, this is a...

pretty specific type of creativity task. It's called the Divergent Association Task. so that's a very specific kind of task that's used in research a lot. And then they also did language-based text creation. So haiku, story synopsis, flash fiction. And so you're getting a snapshot of sort of language-based creativity.

which I think is partly useful for design work, but also design work is not just about language creativity, right? It's about visual creativity. It's about interactive and embodied creativity. ⁓ And so the question there is how do these systems perform? So for example, I mean, I'm gonna give you your royalty of the day by...

Dan Saffer (28:11)
Right?

Mm-hmm.

Nik (28:26)
mentioning micro interactions, but it'd be interesting to see like how good are these things that coming up with say creative micro interactions on a website design as compared to you or other top animation and motion designers. Now at the same time, we are seeing, for example, I bet you the slash animate command in impeccable that we talked about a couple of weeks ago probably gives you pretty reasonable.

motion design, you get some pretty nice little micro interactions. Are they the best? Are they the most creative? Maybe not, but they're probably average. And I bet you we would see similar things here. The question is, is there a point where it starts to overcome? Or is there always going to be this gap between really the top creative people and what

language model-based AI at least, can accomplish.

Dan Saffer (29:16)
I think this gets back to the conversation we had last week all about Demis and Dario talking at Davos was this idea of AI is really great at finishing things that have been set up, but it's not great at coming up with new things that the world has never seen.

particularly in creative fields. And I think that that is the area that probably that 10 % really works in is that idea of here's something we've never tried before. Here's something we haven't seen or let's break this convention Let's come up with a whole different way of

doing music or doing visual art or movies or whatever it is. And I think that that's where that top 10 % are going to really flourish and maybe always flourish. I think this is the goal of AI super intelligence where it can do all those things. It's not the AI general intelligence. It's like a super intelligence that is more creative than any

any living person.

I don't think we're anywhere near that, thank God. And I hope I'm dead before that happens. unless it's like doing things like solving cancer and stuff like that, I'm okay with that. But coming up with new kinds of art and stuff, boy, I really want that to be a human activity.

Nik (30:44)
I guess, mean, I mean.

I would, know we're a podcast for designers, but you know, if there were any doctors or researchers listening, they might say like, no, I want to be the one to figure out how to solve cancer. I mean, why, can't someone else, you know, say, well, yeah, man, I just wish I could have great design so that I can make my product ideas and just move forward. Like, why, why wouldn't I also want that?

Dan Saffer (31:05)
Well, you're going to be able to get that, right? You're going to be able to get very good general design work. Most people don't need to like push the envelope on design or tackle these incredibly difficult design problems. And that's where you hopefully you'll still have a human being involved to help shape that same deal with. I mean, most doctors aren't cancer researchers.

But I would think that for them having tools to find new places to look in,

would be valuable and maybe it's maybe that's also true in design having having something come up with a new way of designing or building something maybe that maybe that is something that would be interesting

Nik (31:46)
One of the use cases that I see AI being particularly valuable for is having this sort of infinite generative mood board, infinite generative content creation to give me visual stimuli. Because what am I doing typically, right, when I'm doing visual design, I want to have lots of input. I'm looking for references and resources out in the world.

I've heard some people say, you know, I kind of miss like the old models that would actually be less precise about things, but they would create more weird stuff, like basically more hallucinations. So for example, in our contemporary image generators,

Everything is super high fidelity. Everything like looks, very realistic, which I think that's what the companies are interested in. That's what a lot of people are interested in. But actually as a designer, I'm not totally interested because I usually have some element of vision in my head or I'm going to be creating a vision through the action of sketching, the action of drawing, the action of modeling on a computer. And I think that way.

But actually just seeing a of weird stuff that even doesn't make sense can be really valuable to me.

Dan Saffer (32:56)
you're longing for the six fingers on every hand someone the other day was lamenting that you now don't get that weird AI text where it's kind of symbols and kind of words so I'm sure that If it hasn't already that's gotta have spawned a whole like art form

Maybe there'll be a retro AI art movement Remember the early ⁓ 2020s. yeah. Well, let's move into our last topic, which is kind of the first in probably what's gonna be an ongoing series around

different UX roles. And one of the ones that we wanted to talk about was the role of user researchers in this new AI world. I think one of the really catastrophic things has happened over the last couple years is the decline of user research. User researchers were hit so much harder than almost anything else.

in the UX layoffs that have happened over the last couple years. And I think we're at a time where we need user research the most. We have a new technology that has very unusual capabilities. Most people don't exactly know how to use it. They don't know the limitations of it.

We haven't really pushed the limits of what we could do with the technology right now. Like if we froze AI technology right now, we would still have years of things that we could be building with the technology that we currently have. And we don't know what the human response to it is. We don't know anything about foundational things like

privacy and security concerns and what people think about AI, what they don't have, things like trust and explainability. We're just so nascent on all those things and it feels like that is a space for user research to really go in and help us define some of these areas, some of these boundaries, some of what

people understand and know and want to do with this it feels like a whole profession has been kind of shunted aside. That's my sense of it. What's your sense Nick?

Nik (35:31)
But Dan, I have synthetic users now. I can...

Dan Saffer (35:35)
Right, can't we just use those and who needs user research anymore? We can just talk right to these synthetic users.

Nik (35:42)
I agree that we need more user research. But again, you and I are probably a little biased here. But I do think that there are some really interesting opportunities in new ways that AI can be potentially used for user research, that it can be used to amplify the work that

I guess maybe now the fewer user researchers who are doing stuff. For example, there is the whole synthetic user thing and there are all these services now that are promoting this. They're building out these synthetic AI agents that act like people. And I think on the one hand, that sounds crazy, especially if you've come up in a world where

You you went out and talked to people, observed real people, but we've had, for example, systems to do text analysis on Reddit posts for a long time. I actually find that asking an AI summarize things on Reddit for me, because, it's probably seen and has been trained on a lot of Reddit data pretty good. And then it can link back to the original Reddit posts and I can go read what people were talking about.

Dan Saffer (36:46)
Don't you worry a little bit about things like...

I've heard it referred to as like safety blanket research where the AI just confirms what the team already believes because it's talking to synthetic people and all those synthetic people have been trained on places like Reddit where they're like mostly like Western educated rich. You know, the basically the kinds of people who are

building a lot of these systems. so I worry there's this kind of confirmation bias that is happening where you don't get the kind of broad and weird and unusual research findings that you did going out in the field.

Nik (37:34)
I think you raise a good point, but a question, how many times have you seen in industry a user research team go out, do great work, and then a VP says, yeah, we're doing it this way, or this is what I want. I think you're right that it's very possible when you're working with these systems.

that they'll give you what you want. That's the sycophancy problem. But if you're going to work with these systems, and if you're using them, do we even have really strong strategies for getting, oddly enough, unbiased AI answers? Because it might be really easy for the for you to even say things in such a way, ask things in such a way, in the same way when we're talking to people that they give us the answers we want.

Are the same techniques, do the same techniques we have for human user research, do they work with the AI or do do the AI somehow start to infer and stuff and then all of a sudden, it starts giving you the answers that you want.

Dan Saffer (38:28)
Yeah, that's an interesting question that I do not know the answer to, so we need to talk with maybe some actual user researchers. Let's have some guests on the show.

Certainly when you're out in the field, you can have these exact same problems where research subjects don't want to tell you like, this is a really bad idea because they don't want to hurt your feelings. Maybe in a way that an AI synthetic persona is not going to do that. They may not sugarcoat it. If you can tell it, hey, don't sugarcoat this to me. It might actually obey those instructions in a way that human beings might not.

Nik (39:09)
one of the other potential benefits of

AI and its incorporation into user research is people may be more willing to do certain types of user research and they might even be able to do at least average user research. So for example, synthetic user stuff, mean, we can we can debate is it good? Is it not good or whatever? But if that's something where a company like a startup is like, well, we really don't have enough money to do a proper

user test or really to go out. So we're just going to do it quickly with this. I mean, it might be better than anything they had ever done before. Things like user testing. mean, this is something I'm really excited about, right? This idea that you can have, say AI proctors to have someone do like, Hey, let's do a contextual inquiry session. And it's just an AI facilitated contextual inquiry with a screen share. Like I showed students how to do this in my class, where you can fire up.

Google AI Studio Live mode, you enable the mic, you enable the screen share, you throw in a system prompt that has maybe your contextual interview guide, you put all your rules, you put all this stuff that you've really thought about, like this is how I administer this kind of stuff. And then you let it do it and it's like, ⁓ it's pretty good. It actually works reasonably well. And the ability to say scale that, now get that out to a hundred people. I mean,

Anthropic actually did this. had an AI interviewer in Claude so that they could interview people on how they were using AI. And so their team was able to do this pretty quickly. And then of course they also used Claude to sort of analyze the data. And there's been some hot takes on this and things like that. We can link, I'll maybe try to find some articles. I'll link it in the show notes. But you know, people are saying, well, it's really not getting in. If you actually look at the...

transcripts that people wrote, there's a lot more richness in there. And that may be true, but again, this ability to go out, get like a thousand people and get that data is something we've really not been able to do at scale before. It's something we've not been able to do with the cost that we can do it now. So there's lots of potential positives here. at the end of the day, the goal of user research

is to hopefully help you make good decisions in your design process to then lead to great products. But the measure of a great product isn't always based in how great was the research that led to it. That's the thing that helps us get there. It's still something that plays out in the market.

Dan Saffer (41:38)
I also think when we're talking about agents, how much of understanding workflow before we start to add AI to it is a user research task.

What is really important? Where are the key decision points in here? Doing things like task analysis. Those are all what I would consider user research and understanding like this is a great place for AI to intervene. This is a terrible place for AI to make this decision. I feel like that is a user research task, whether that's done by designers or user researchers. think understanding those kinds of contexts

I think is going to be super important as we move into a more agentic world where we're just throwing AI at all different kinds of workplaces and processes. We first need to understand the processes and we need to understand the workflows before we start disrupting

Nik (42:41)
Yeah, I agree. And I think this is not the only time we're going to talk about user research on this show. I imagine we're going to come back to this topic because I also imagine that in six months from now, there's going to be new tools out. There's going to be new people trying stuff. I mean, there's new research coming out all the time on how people are using it for user research. How good is it at qualitative analysis? How good is it at representing people in a synthetic way?

Dan Saffer (43:11)
Absolutely. And that's our show for this week. We'll be back next week with more news stories, more hot topics, and maybe even a special guest. We'll see you next time on AI and Design.