Dan: Welcome to AI and Design, where we explore how artificial intelligence is reshaping the world of design. I'm Dan Saffer, and I teach at Carnegie Mellon's Human-Computer Interaction Institute. My co-host, Nik Martelaro, is in attending the biggest conference in the HCI academic world, CHI, and he'll be back next week. with some of the latest research being done in AI and design. So it's just me this week. Your boy Dan, oh Danny boy, me, me, me, me, me. It's the Dan Saffer experience now. But on today's episode, I'll be giving some thoughts on some recent AI and design news, just me. And what is that? Well, Let's dive into it. So Jenny Wen has gotten a lot of buzz, but her boss at Anthropic, Joel Lewenstein, head of design, has some things to say too in a new interview that is surprisingly hopeful. And hey, we can all use a little hope right about now, right? But then Dimitri Kargaev puts Anthropic and its ilk on Blast, claiming that AI products have terrible UX and he thinks he knows why. Then I'll look into the aesthetics of AI, particularly focused on AI brands. then we'll end with a call to arms from Patrizia Bertini on what the future of human centered design should be in the age of AI. Not much technical this week with Nik not being here to explain it to me. So we just got some good old fashioned design stuff. So let's get into it. Fast Company released an interview last week of Anthropics Head of Design, Joel Lewenstein. And I was surprised ⁓ this didn't get more traction because I thought he had a bunch of ⁓ things to say. The first one, which we've talked about on the show a lot is that friction can be a good thing. And of the things he said was, yeah, we want you to go back and ask questions and have Claude prompt and ask questions. So I thought that was a pretty interesting thing. Really this idea of cognitive forcing functions of forcing people. to really go back and people think about the decisions choices they're made and just not run off and do it for you. And he talks about Claude and I quote here, ⁓ should be a sparring partner with you, which I thought a kind of a radical take about fighting sycophancy and really pushing back on users. And that's also something we discussed on this show quite a bit, especially back when we had Chris Nossel on to talk about assistive technology. one of the things that I liked about what ⁓ Joel was saying here was that by design has opinions and designing ⁓ the breaking point those opinions is actually really interesting. And they kind of get into this a lot in this interview. And near the end and they talk about the kind of personality traits that Claude or any AI can or should have. how much should it look after your well being? And one of the examples they give is like should Claude say things like hey, it's ⁓ 2 AM. You should be going to bed right now and of hosts is like yeah, that seems a little paternalistic to me and I have to agree like I wouldn't want Claude. giving me those kinds of instructions and looking out for me in quite that kind of way. Like, you know, stay in your lane Claude. If you're gonna have comments about what we're talking about, fine. But if you, as soon as you start to bring in kind of outside context that you know nothing about, maybe that's the point to like pump the brakes there. But I think it's a really interesting place for designers to be in right now where we not just designing the thing, but also having to think about the personality of the thing. We didn't have to think about designing the personality of spreadsheets or ⁓ social media tools or applications. Those things kind of left to brand to kind of bring some of that in. But our job, particularly in the interaction design world, was to make things very frictionless and easy to use and not as much about having opinions, having personality. And so I think that is an interesting space that a lot of us are moving into. Some of the space that certainly conversational designers have been doing over many years. In the interview, Joel talks a bunch about organizational power. And this is definitely something that's been on my mind lately, especially after conversation last week with Mike Kuniavski and plug here for my own stuff. wrote an article about it called, we're all doing the same job right now, which I'll link to in the show notes, but he talks about how Anthropic works. And one of the things that he says ⁓ makes the prototype really drives the decision-making ideation and roadmap. the people coming in with the prototype, they have a lot of controls. So I wonder if that really behooves us as designers to really get in front of and it often a into places we frequently have not been able to go, which is places like, into product where say, well, this is the next feature on the roadmap. Let's build that. This is a time where. designers have to step up to say, I think that this should be the next feature in the roadmap and take control of that. Joel talks about that this really happens at beginning of the process, ⁓ that does this is really in charge at the beginning. And then eventually that there is a place for all three roles of PM designer and engineer. But at the beginning, you do have this role collapse, and then eventually it splits off into individual roles. So, at Anthropic, and Joel definitely a place for all three of those roles working together, and not just a ⁓ complete role collapse. And at top of the I promised you some hope, here it is. they are working now, like probably many organizations are going to work in a year or so or in two years, and they are doubling their design staff. They are really leaning into the value of design, really into the value that the three different roles are bringing to the table. And I think that's a really sign. because that was not the case a couple of years ago at Anthropic or seemingly a lot of these other AI companies where it was very heavily engineering and research driven. And we'll get into that in our next segment about ⁓ why UX so terrible at some of these AI companies. But ⁓ I think is a space to be And I really think that a lot of companies start to realize that when you start adding in AI into your products and services or building complete AI products and services that you need more design work, not less. Even though the builds can be faster, you still need some of that design thinking. He talks a lot in this interview about the biggest challenge across all their design teams is that the models have gotten so much better. But now there's a real question about how do you align the power of the models with the user experience? And he says, how do get the power of the models to be useful to you? And I think that's a really interesting area of disconnect. How do you start to... surface value and power of these models in way that makes sense for both power users and for your users. And at one point in conversation, he starts talking about finding the right level of abstraction for different kinds of users. And I think that's a really important point for designing for AI. how much detail are you really showing? example that they talk about is that who aren't power users, they don't care that there are maybe 10 different agents working in the background doing something just that it gets done. Whereas if you're a very technical person, yes, you may actually want to look at all of those agents, look at what they're doing, see where the blockages are, see if you can fine tune them. you are on top of them almost as though you are on top of a team of people managing their day-to-day stuff where casual users or just less power users don't need to look under the hood. You don't need to see that. And so I think designing for AI often means thinking about those different levels of abstraction. Now I will say that one thing in the interview that made me go, is over-reliance ⁓ language and prompts to get users to where they need to go. And I think because Anthropic, and probably most AI they're very red-pilled because they're power prompters, conversation for just becomes the default UI for everything. sure, sure, you know, using conversational to get started great, but eventually you have to expand your toolbox back to using some of the graphical user interface paradigms that we built up over the last 40 years because those are still valuable. Those can still be better in many cases than having to describe what it is that you are trying to do in language. In the interview, they're surprised that people don't write these long prompts and it's like, well, yeah, we've been trained for 25 years to write these very short prompts and short texts. And people just don't write these super long prompts because that's just not what we're used to. And a lot of people are not very ⁓ or writing-based. And so ⁓ think that is a flaw I think that they need to start thinking about. And we in general just need to start thinking about ⁓ when the moments that are good to start building in. graphical user interfaces, getting the stuff that we know back into the hands of users, generating it on the fly in GenUI to users do the tasks that they're trying to do with the tools that they know. Okay, which is a good segue into our next story. So over on Hack Noon, Dimitri ⁓ Kargave, who's writing as D-Flect, has an article called, AI products have terrible UX. Here's why. His thesis is basically what my HCII colleague, John Zimmerman calls the AI innovation gap. And it goes something like this. people who understand things like how transformers work, they don't think about UX. They're not thinking about flows. They're thinking in loss functions and benchmark scores ⁓ those kinds of technical details. So Demetri says here, you know, when a machine learning engineer ships a UI, you get a settings panel with 40 sliders and tooltips that say things like control the randomness of model output. And that's not particularly helpful. On the other side, you get designers who understand interaction design, but ⁓ know anything about AI systems. We can do user journeys. We can... do lots of GUIs but we don't know things like what a context window is. And we don't enough about trade-offs. And we oftentimes just treat AI as this black box that works or it doesn't work. And this is what John Zimmerman calls the AI innovation gap. And that you've got these two groups who don't speak each other's how do we start to bridge that gap? Because the intersection of people that can do both is genuinely small. And that means that a lot of AI products, they get built by one camp or the other, and you really start to feel it. And he starts to outline some of the issues that this brings up. And it's hard to argue with a lot of these. So the first one is, Chat for everything syndrome. this is what we were just talking about with Anthropic is that developers like use chat box as this ⁓ lazy catchall that, a tool has a specific purpose like summarizing, it probably should have a button or UI element then force ⁓ users to type ⁓ the like summarize this for me. The next thing that he... that he flags is poor loading states. is slow, and simple kind of spinning wheel or just nothing just isn't enough. And how ⁓ we start include things like ⁓ better progress or streaming text in order to indicate what the AI is doing and respect the user's time and attention? As a side note here, what do we do when you have these tasks that are incredibly long? That is a real unexplored area in AI and design research or just AI and design out in the world. He goes on to talk about bad UX being things like exposing a lot of technical settings like temperature or top P or strength or those kinds of things and these are all things that are too technical for most people to understand. Those things should be behind an advanced setting some kind of progressive disclosure that happens over time. The next issue that he calls out is treating output as final, that all AI results should be treated as drafts and good UX allows users to go and edit and regenerate or reject specific parts of the output in line, in context. that alone makes the AI feel more like a collaborator rather than a black box you send something to and eventually get something back from. And the last one he talks about is this idea of these missing feedback loops many had things like thumbs up or thumbs down. If they have that at ⁓ all, just like fix this button and these of things really prevent the system from learning from user corrections. So how do we start to better incorporate those? Even simple things like thumbs up and thumbs down are good, but even better, more specifics about what is wrong and why. Last thing I want to talk about with this article is a claim he makes kind of early in the article. it stuck out to me and I've been pondering ⁓ its veracity ever since I it. And it's this, and he says, mediocre model with great UX beats a great model with terrible UX every time. And I'm like, hmm. Is that true? And the example he uses is the original chat GBT where it was simple. It let people kind of start to do it without having any kind of technical sliders or any of those kinds of things around it. And people started to explore with it. And it took off and that leads us to where we are right now. Now the question for me is sure that was true. three years ago, but is it really true now? Would a mediocre model with really great user experience beat out a really state of the art model that has terrible user experience? I wish Nik was here to talk this through with me, but I'm a bit dubious of this claim. think that once, especially people who have sampled the better models, I don't think a mediocre image output model with great UX is going to beat a better image generation model with terrible UX. Certainly not for power users, maybe for regular users or people that don't need things to be professional that they aren't using for work. But ⁓ am very ⁓ of this claim, but I'd love to have a discussion about it. Listeners, let us know. Let me know what you think about this claim from Dimitri. Let's go on to our third piece, which is about AI branding. The design studio, A Color Bright, they did a brand study called the Aesthetics of AI. And rather than look at things like the UX flow or the model output, were simply looking at the branding and brand identity of some of the leading AI products there. Because so many of these products do the same thing. So the branding and identity is one way to set them apart. And maybe now that I'm saying this, maybe this ties into the article we just discussed where maybe a well-branded product would perform better than a poorly branded or clunky looking AI product. This is certainly an old HCI because Don Norman years ago said beautiful things better. And I think maybe that here. But that's a side note. So they came up with five different categories. So the first category was ⁓ likable leaders. ⁓ And these brands that for mass market trust and quiet luxury. They use soft and offensive aesthetics to signal that they are the responsible adult voices in the road. So in this bucket, you have things like Anthropic, you have Microsoft AI and you have Gemini. These are things that use muted tones, beiges, elegant typography, shades of off-white, designed to be extremely ⁓ neutral Things like Microsoft AI, employ these very soft gradients very kind of approachable AI that really blends into corporate ⁓ which because it's things Microsoft AI and Gemini, they have to blend into their suite of things right now. Anthropic, of course, being the outlier here, you know, using much more muted tones, more beiges. but they all fall into these likable leader category. second category they talked about were the gentle humanists. And these companies position AI ⁓ a partner to human creativity rather than a replacement. And the branding for these is often kind of low tech. It's more organic in feel. These are things like Notion ⁓ or ⁓ things like notion, have things like hand-drawn sketches and human centric illustrations to make the notion AI feel more like a tool for thinking rather than something generated by a machine pie uses things like soft readable fonts and a very kind of minimalist emotional tone. ⁓ That's focused on conversation. You don't see really any of the. major players in the gentle humanist category. But you do see some in the category, which is nerdy idealists. ⁓ this group prioritizes technology and developer community ⁓ polished corporate branding. The look is often intentionally very kind of unbranded or quirky. OpenAI is kind of the poster child here. So while they do have the kind of likeable leader qualities, they're often leaning into kind of non-brand, academia, aesthetic, kind of research papery, uses very technical, slightly unpolished visuals like the swirl logo. But you also have other things like Mistral AI, which has a distinct logo and of branding that appeals to more kind of open source developers and European technical products. The fourth category is bold builders. And brands, lean into kind of more of the power and vibe of AI. This is all the technology and they frame it as this kind of vast frontier crossing force. And poster child here ⁓ is our friend Mecha Hitler. AKA XAI. Kind of the definitive ⁓ builder, because the branding is dark, it's high contrast, it's industrial. The logo uses geometric black shapes. Nothing nothing organic, no off white beige here. No pastel gradients. It's more about technological order, precise computation. Perplexity is also similar. It uses these kind of dark modes and cosmic imagery. They have these kind of sleek futuristic search interfaces it is all about this powerful alternative to what else is out. And the last the categories that they found were these companies they categorized as utopian dreamers. And they kind of create entire worlds and use kind of surreal or retro futuristic visuals to suggest that AI is going to unlock entirely new realities. And these are places like mid journey or runway or world labs. just ⁓ different kinds of feels than any of the other categories. you'll note that none of big players really seem to be in this category. So I suggest taking a look at this article because it's got a lot of really cool visuals. It's got a lot of ⁓ brand dissections of each of these different categories of company. And it's interesting if you're thinking about how are going to position yourself ⁓ startup, for instance, ⁓ might want to think about moving into categories that are not the likable leaders, because that category might be oversaturated and you might want to come off more as gentle humanist or nerdy idealist, or maybe even as bold builder or. utopian dreamer just to really distinguish yourself and set your product apart. Let's wrap up on what I think is a really thoughtful piece Human-Centered Design Has Grown Up. It's Time We Did Too. I think her subtitle really spells out her thesis pretty well. she this, says, we spent 25 years making technology work for users. Now we need to make it work for human beings. These are not the same thing. And the distance between them is where a generation of harm accumulated. And the of this article ⁓ is that had this traditional era of ⁓ human centered design. And she started 25 years ago, I would of course, probably started, you know, 30, 40 years before that, but let's go with it. So she says that, around 25 years ago, you had kind of the traditional era or maybe the golden age of human centered design. And that focused primarily on usability and removing friction. And that is no longer sufficient at all has become in some ways complicit in systemic issues that we're seeing right now ⁓ in attention economy and in digital manipulation. And ⁓ is true now in the AI She starts with Steve Krug's Don't Make Me Think, which was a classic book that came out around the year 2000. then also references our old pal Don Norman. and their goal was to make technology respect human cognitive reality. she says, and I think rightly so, that this succeeded really brilliantly. We have done an amazing job the last couple of decades removing And our have become great. They've become really intuitive, at least ⁓ up until But that's a topic for another time. By removing all friction, designers have unintentionally built the infrastructure for things like the attention economy where users are managed rather than empowered. need this shift from thinking about ⁓ users to thinking citizens. In the old way of thinking about things success was by task completion, conversion rates, and our pal NPS. And these measured transaction. Was the transaction complete? In HCI, we used to do these old GOMS studies. How many clicks ⁓ and how long it take before the transaction was complete? In a new reality, design must now consider whether the system leaves the person more or less autonomous. measures the ⁓ of people rather a transactional view. What does that mean? So that means that we need to stop viewing people merely as ⁓ users, aka of a product. and start seeing them as individuals with fundamental rights, aka citizens. one of the ways that Bertini says that we should start doing this is Learn how to read regulations as though they are a design brief. We need to be able to look at translating legal requirements into interaction requirements. And the example she gives is all around the EU AI Act. that's all about manipulation, deceptive AI systems ⁓ illegal and regulations that are now focusing on things like data and decisions and accountability. We need to start getting deeper into understanding those so that we can start to reflect those back in our design practice. And so the design question is do we move beyond this old narrow question of usability, which is can the user complete the task to this new question of does completing the task leave the human being more or less free? And I really like this. I wrote an article a couple of weeks ago about existential design, and I think this fits right in with that. I'll put that link in the show notes too, but Her advocacy that human centered design has grown up and we need to grow up with it because now the stakes have changed. It's no longer just about making things easy to use. It's about protecting human agency and ensuring that ethical behavior ⁓ that path of least resistance. And need to ⁓ start a world where technology respects the fundamental rights of the people using it. And can't recommend this article enough. I think it's a great way to frame the issues that are facing us ⁓ designers these very complex, potentially dehumanizing and really think that this is worth a lot of discussion and a lot of thinking, both in design curriculum where I work and out in the world in professional practice. read this and start think about making users ⁓ into citizens. And that the solo show for today. Wow, was harder than I thought talking about these things by myself for so long. Nik, I didn't realize how much I missed you here, my friend. I'm really looking forward to that. And I hope you are too. We'll see you next week here on AI and Design Podcast. Be sure to like and subscribe and all those good things. Bye bye.