An AI talk I once gave

We’re in the AI age now. I’m not sure where exactly in the age we are as there’s a lot left to play out, but I can confidently say we’re near the beginning.

We were even earlier in the AI age in 2024 when I was asked to speak on AI at a design conference in Singapore, which I was happy to do. Below is the transcript from my talk. This was given orally, so lacks the polish of the written word. And it’s been only two years since this talk, but right now a year feels like a decade, and I’m glad it kinda-mostly holds up—though reading now I can spot areas where my thinking has matured and expanded. Which is to also to say that I have a lot more to say on AI, some headier and less practical than what’s here. I’ll give more armchair-philosophical as time allows.

Hello everyone, My name is Kris Bell and it’s great to be here today–thanks for having me! So you know who I am, a bit about me: I have been in UX for a while—I built my first webpage in 1996 and I’ve been in the user experience industry for over 20 years, starting in the early 2000s designing and building websites for small business and not-for-profit entities. I’ve worked in consulting, at Google, for a brief time at a startup called Flexport. I’m currently a Director of UX at Amazon in Seattle, where I’ve spent, across a few stints, nearly 14 years. I should also mention that even though I work at Amazon, I’m here today not as am employee, but as a citizen of the world. So my comments are mine, and mine alone.

Having been in tech as long as I have, I’ve watched as a whole bunch of different technological innovations have rippled through the industry. I remember when Web 2.0 and Ajax was a thing that everybody was trying to sprinkle into their applications. I rode the Mobile and Responsive Design wave, Design Systems, Block Chain, AR/VR, NFTs … and now AI. In case you didn’t notice, AI is having a moment right now.

AI is obviously a powerful suite of technologies. I’m definitely not an expert in AI — far from it. I can’t predict what impact it will have on design specifically or tech broadly, but I do think that there will be applications of artificial intelligence that will fundamentally alter behaviors in our world. I also think it’s highly unlikely that the use cases that different folks are chasing right now will be the ones that are enduring breakouts. Time will tell.

So I’m no expert in AI, but I know a thing or two about design, so I’m glad I get to focus on that. There are a number of ways that we could think about this. I’m sure a number of you have given a lot of thought already to how we could use AI within your design workflows to generate layouts or illustrations or graphics, maybe write some good marketing copy, or some other way to shortcut a lot of the craft we put into our work. I actually don’t want to talk about that.

Rather, I want to talk about designing AI into our actual applications, point out some of the pitfalls I see, and hopefully provide some tips or pointers on how perhaps we as an industry ought to be thinking about this.

I think I’ll start with what I think is one of my most important points first. And I think this is obvious, but needs to be stated for the record. Which is: AI is not an end unto itself. It is a technology, and a technology’s value is zero until the point at which it can be applied in a useful and meaningful way to generate value. As designers we don’t work backward from what technology can offer. Rather we work backward from what we know our users’ desires, hopes, fears, or other motivations are. And then we pattern match. How do we characterize the goals of our users and then, what available technologies can best meet those goals? So, despite the promise AI brings as a panacea for inference- and sense-making out of sea of data—just remember: it’s a tool in the toolbox.

To put this slightly differently.

Out on LinkedIn, on medium, on Reddit, and a whole bunch of other places I’ve read articles with titles such as “Designing for AI” or how to design AI into your application. And I think this is simply fundamentally the wrong framing. When you think that way you get products like the Humane pin or the Rabbit R1 which seemed to ignore some of the same basic experiential questions that have plagued all voice assistance since they launched—things like “do people want to talk awkwardly to a computer in front of other humans?” and what truly is the advantage of voice + ears over touch+screen — other than having to dig through your pockets, of course. And not to insult Humane or Rabbit too much—lord know they have been ripped to shreds out in the media enough. So, the lesson here, of course, is always solve for humans first.

Now, obviously you don’t want to ignore the power that AI can bring. So a good frame to start with instead of “what can AI do” is “what problems do I need to solve, where might AI provide a good solution?” Or “what problems have we solved already, but which AI can improve upon?”

Once you’ve done that, then maybe you’re ready to get into actually working through the nuance of the technology. Which means it’s design time. So, some thoughts on that.

I said I wasn’t going to talk about workflows. I lied—I have one. And that is, aspects of your job might need to change. Why? Well, generative AI is only as good as the inputs it’s fed. We’ve seen some of the gaffes that Google has faced. Have folks heard about the best way to keep your cheese on your pizza? Glue, of course! In this instance, it wasn’t hallucination that caused Google Search to spit out a bonkers answer. Rather, it was the fact that Google crawled Reddit, and Google’s model couldn’t distinguish between a sincere answer and a joke. Often time as UXers, or as PMs, we leave what’s happening under the hood to our engineering partners. But we’re the ones responsible for what is presented to users, and how it is presented—the visual artifacts, the structure and hierarchy of apps and pages, and the content. And in the case of Generative AI— what is shown to customers is completely dependent upon training data, so as those responsible for that experience, you have to be involved with what is fed in – for the purposes of quality, accuracy, and inclusion. In a sense, you become a curator.

What might this look like in practice?

Say, for example, you’re designing a navigational menu for a library app. And you want imagery for each node in that menu that can scale to some large number of use cases. In this case, the training data used is hugely important, and the designer ought to be the curator of this particular style, feeding the model very specific images that meet the criteria for the visual bar set, so that the model generates a high quality, consistent set of images. Or let’s say you’re working on a government procurement website with lots of big open text fields that are meant to take in thousands of words for an RFP (an RFP, for those who don’t know is a Request for Proposal — essentially a business or government puts an RFP out to let business know they can bid on a contract). Often RFPs have very specific criteria about what is submitted. And they want lots of bids, and for them all to be well written, so they can get the best quality for the price. If both the rules for the RFP and examples of well-written RFPs were fed to a writing assistant that offered contextual help in real time, it could help RFP authors be far more successful than generic writing tips would be. And positively changing user behavior is job number for designers.

The RFP example leads to another insight – which is, if you’re applying artificial intelligence to an existing application, you want to consider deep how it integrates….As the land rush for AI speeds up, we see examples out in the world, where a company’s AI application is just bolted on to their existing interface. A good example is the continued rush towards chat bots. In a lot of applications, the chat bot is represented as a floating action button somewhere on the screen. You’re required to tap it, formulate your query, submit that query, and get a response back. Now, in some cases, the posture of this “AI Assistant” doesn’t help the user understand why they might use it. For example, Bing presents its Co-pilot alongside its main search field. They both take queries and output results—essentially, to a user, they’re kind of the same thing. Google, on the other hand, integrates AI right into its search interface, and makes a determination based on the nature of the query, what to serve you. Assuming for a second that the models are the same – what is a better experience?

Of course, in my view, you should deeply integrate your AI application within the existing interaction paradigms that you’ve already defined within your application. Otherwise, you’re teaching users two ways to do the similar thing, and potentially ruin the efficacy of both. Want to develop a writing assistant for your RFP application? Evolve the authoring interface to provide writing guidance within context of the text box, rather than as suggestions in the sidebar. Native-AI applications are a different story, but I have a feeling we’ll need to straddle the line between conventional UI and AI interactions until the point at which most folks in the world have become familiar with whatever new interaction paradigms emerge.

OK, so you’ve found that AI is the right technology for a problem. You’ve helped train it, and think the results are good. You’ve integrated into your products.

In order to test AI-powered applications you’ll often need to use naturalistic experiences. This is because one of the promises of AI is deep personalization based on either latent inputs—something that it knows about you—or based on your specific prompt. For generative AI applications specifically, it’s highly unlikely that you’re going to be able to arrive at any sort of conclusive insights using a Figma prototype when the quality or efficacy of the output is judged relative to what the particular user was after. I think this necessarily means that you’ll need to build a working prototype and work with the scientists who are building or adapting the models. In a lot of ways, this means that the product development experiences are no longer a three-legged stool, but a four-legged chair.

And as long as we’re talking out user perception, you need to make a choice whether to tell users it’s AI or not. My basic rule of thumb is this: If your application of AI has no right or wrong answer, you probably don’t need to label it as AI. Examples might be a ranking algorithm that re-orders the UI. Meta doesn’t need to tell everyone that Facebook and Instagram use AI to rank your feed and surface recommendations. If Microsoft retrains its spelling and grammar checker on a new set of updated grammar rules, and it shows a 20% improvement in terms of accuracy, they dont need to tell you. In the end, both the ordering of feed, and suggestions for better grammar are entirely in the hands of you the user, and neither applications is generating something new and presenting it as fact. Rather, they present as a set of suggestions, which you can either take or leave.

However, if you’re using generative AI to summarize a body of content or to generate something new, I think it’s incumbent upon you to let users know that this is artificial intelligence. And the reason for that is that the user needs to understand that there is a risk of inaccuracies or hallucinations. To put it another way, if we compare summarization with Facebook algorithm for a second period the algorithm example is self evidently subjective. It’s obvious on its face that there is no organizing principle behind the feed. But summarization, image generation, etc. comes with a patina of objectivity—factuality—and this is why I think it must be labeled as AI generated.

So, you’ve found that AI is the right technology for a problem. You’ve helped train it, and think the results are good. You’ve integrated into your products. You’ve tested it. All is looking good.

But you have another question to answer. Is it worth it?

In years past, I never would have mentioned the cost of a particular technology as applied to a design solution. But with generative AI, I believe the designers have to be far more aware of the actual cost. And I’m talking about at least three costs here: one I think is the actual cost of the infastructure needed to deploy an application that has generative AI as part of it technology stack. This is probably a point-in-time concern, but the cost of compute right now is so extraordinarily high, that you have to be aware of both of the cost and then what you hope to get out of it. Because if you don’t, the risk you run is not shipping an application at all because there’s no way you’re gonna get a return on the investment. So you’re creating a revenue-generating application such as a marketplace for selling hotel rooms, you’ll want to develop a hypothesis about how much some AI-powered feature will increase incremental revenue. Or if you are summarizing a knowledge base to make customer service agents more efficient, you’ll want to understand how much more efficient it will make them, and whether that cost savings is more than simply hiring more folks.

Another cost, of course, is the cost to society and the world. Generative AI is extraordinarily expensive because of the amount of electricity required to power that compute infrastructure, so there is a real environmental cost to generative AI. According to Morgan Stanley, by 2026, global AI-related electricity consumption will increase to be roughly equivalent to the energy consumption of Sweden or Germany. That’s massive.

Finally, a third cost is the cost to your brand. There’s a land rush because AI right now has a brand-burnishing effect on brands, kind of like having a website in the 1990s was brand burnishing. So as your and your team think about your use of AI, are you doing so in part because it will look better because of the optics? And if so, how do you weigh that against the cost to the environment or the financial cost of compute? And how also are you factoring in generative AI hallucinations, which at best can be somewhat embarrassing, but at worse can be harmful to society and the world because of the spread of misinformation or the psychological effect on users?

These aren’t things to take lightly, and I think it’s incumbent upon every company to develop an ethical framework that teams should operate within, and I think designers and product managers should be the leaders in the creation of this framework. Otherwise, you risk operating as mercenaries without a code of conduct.

That’s a bit of a dire note to leave you on, so I’ll end with this—the same point I started with. We’re designers. AI promises to be a transformative technology. As you apply AI more and more to your applications, do so in service of the humans you serve. Ensure that what you’re doing is beneficial to them. It’s what we in design have been trained to do, and its what we’ll continue to do once AI becomes a boring part of the technology stack that we take for granted every day. Thank you.