The Neuroscience of Attention in AI-Driven UX

Julianna Goodman at UXDX Community: Attention, Trust and Intelligent Brand Experiences. Video: https://youtu.be/UuU4PtWb1Qg

Readable transcript: edited from the recording's captions for readability (fillers and false starts removed, punctuation and section headings added). Wording is the speaker's own. Timestamps are positions in the video. Names marked [?] could not be verified against the audio.

Why people hesitate before a simple decision

[00:00:08] Hello everybody. Good afternoon. I hope everyone is doing well, and thank you so much for being here today. I know we all have a lot on our mind, so let's ease in with something very simple. Let me start with a question. Why do people sometimes hesitate before making a simple decision? It sounds small, but it happens all the time. Let me give you an example. Have you ever sat down at a restaurant table, opened up a menu and realized that there were just too many options? You scan it once, then again, and then reread the first thing you saw again, and after a minute you are not excited anymore. You're tired. So what do you do? You order the same thing you always do, again.

[00:01:02] This tiny moment of hesitation, that pause before we decide, is actually where design lives, and it is something that I am very passionate about and spend a lot of time thinking about. By the way, I'm Julianna Goodman and I lead design strategy in my work. I pay a lot of close attention to those tiny little pauses in our digital experience, the moment where someone hesitates instead of moving forward, because that hesitation doesn't just happen with the restaurant menu. It happens in our digital platforms today. That's what brings us to today's talk.

[00:01:49] Now let me bring this moment into the digital experience. Think about what happens when someone opens up a website today. There is navigation across the top, a bold headline, multiple calls to action, maybe a pop-up appears, maybe a chat starts blinking in the corner. Every element is technically very helpful. I'm sure there are good reasons for putting all of it in there, and none of those things are really wrong, even individually. But everything asks for attention at the same time, and nothing becomes clear.

[00:02:37] Your brain receives 11 bits of information every second through your senses, but consciously you can only process about 40 bits of them. So attention isn't unlimited. It is a scarce resource. Just think about the ratio for a moment. When someone opens your digital experience, the brain immediately starts filtering: what matters, what can be ignored, and what feels safe to click. When there is too much to sort through, people hesitate, they freeze, they abandon the task. And here's an interesting twist. We all think that AI was supposed to come and help us solve this problem, but let's be honest here, many AI interfaces are actually making the problem worse. You have more content, more suggestions and more decisions to make.

[00:03:33] Before we go further, let's quickly share what we will explore today. First we will start with how the brain actually filters information. Second, we'll explore why humans trust some AI systems and hesitate with others; the reaction isn't random, there's a reason behind it. Third, I will share three design principles for working with attention. And finally, I will show you a real copilot project where we applied these ideas.

Attention is a filter, not a spotlight

[00:04:10] Let me give you a simple example of how attention works. I don't know how many of you are parents here, but imagine a parent sleeping through the night. They might have some traffic outside, a door closing somewhere in the house, maybe the heater turning on. All of those sounds are happening at the same time, and yet the parent sleeps through it. Now imagine a baby makes a small cry: suddenly the parent wakes up instantly. What's interesting is that the brain was actually hearing all those sounds I mentioned earlier the entire time, but most of them just got filtered out, because a baby's cry is a signal the brain has learned not to ignore.

[00:05:00] This is where our common understanding of attention is a little bit misleading. We all thought that you describe attention as a spotlight, but neuroscience shows something different. Your brain is constantly filtering information, and only the signals that are meaningful and relevant get through. Good design helps the right signal pass through the filter. This matters a lot for digital design. Attention isn't about what we control, it's something that the brain automatically protects, because every screen we create, every notification, every headline, every AI suggestion is really trying to pass through that filter. That's great, but when everything happens at the same time they start to compete with each other, our brain is overwhelmed, and then we just block everything out. So remember, good design isn't about adding more signal. It is helping the right signal go through the filter.

Trust is biological

[00:06:10] Attention is just part of the story. There is another process happening in our brain at the same time, and that is trust. Trust is not just how we feel it. It's actually biological. There are two parts of our brain doing this process. One part is called the amygdala, constantly scanning for threat and uncertainty, and at the same time another part of our brain, the prefrontal cortex, evaluates what is and isn't reliable. So every interaction goes through this very quick internal check. Your brain asks a question: does it make sense, is it safe to act on? When a system feels clear and consistent the brain relaxes, but when something feels like a black box the brain becomes very cautious.

[00:07:12] That's when the hesitation comes in, when we take a moment to make a decision. Just remember, our brain really doesn't like to act on things that we don't understand. This is exactly the challenge with AI. If people want to trust AI systems, we have to design experiences that reduce that uncertainty.

Three principles for designing with the brain

[00:07:38] Now we understand a bit about how the brain filters information and how trust works. The real question becomes: how do we design AI that works with the brain, not against it? In our project, three practical principles started to become our north star. They help guide how we design AI experiences. First, we reduce cognitive load, because the brain is already filtering a huge amount of information. Second, we build trust through transparency, so the system is understandable rather than a black box. And third, we keep the human in control, because AI should just support people. It should not replace their judgment.

[00:08:35] These principles might sound a little bit simple here, but they become very powerful once we start applying them to a real product. So let me show you how they play out in the copilot project that our team is working on today. This is a web copilot project.

Where should the copilot live

[00:09:00] Before we got into all of that, we started by looking at the help ecosystem on our website, and we noticed something that is so important. Users already have so many ways of getting help. They could search, they could browse documentation, they could read FAQs, and they could even contact us or connect with our live support. Each option was designed to help, but when all of them appear at the same time something else happens: your brain cannot decide, and all your brain is asking is, where should I start? When people are presented with too many entry points the brain doesn't become more efficient, it experiences decision friction. People hesitate, they pause, sometimes they abandon the task altogether.

[00:09:56] Then we faced the real challenge. We were about to introduce the AI copilot into our system, and that is not simply adding something to the page. When we do that, we could actually make the system more confusing than ever. As we explored this problem, we realized something very important: it is not about adding an AI copilot, it is really about where the copilot should live in that experience.

[00:10:37] On the marketing side people are still exploring, they are learning about the possibility. In this context AI can support discovery, but it's not necessarily the main entry. It can be in the corner, at the bottom or on the right-hand side. That's just the standard practice with a lot of our competitors.

[00:11:04] But inside the platform the situation is a little bit different. Users are already working, they are already trying to complete a task, and when someone gets stuck in the middle of a task they usually have a simple question: how do I do this? This moment, when help matters the most, is right here. So placing the copilot directly in the working environment means guidance appears exactly where the question happens. On the marketing side AI supports exploration, but inside the platform it supports execution, and execution is where people need help the most. That led us to a key design question: not just should we add a copilot, but where should it live?

Reducing cognitive load

[00:11:57] The first principle we focused on was reducing cognitive load. Earlier we talked about how the brain works more like a filter than a spotlight. So as designers our job is really to support that filtering process by showing people only what they need and when they need it. If you look at the screen here, when the user asks a question the copilot gives an easy-to-read, clear answer first: just the key steps, nothing extra. Then when the user wants more context they can click on the detail, and now the additional explanation appears. So the interface stays clean and focused while detailed information is still available for someone who actually needs it and wants to get more information.

[00:12:59] We also use visual hierarchy to guide attention, making sure that the answer is the most prominent thing on the screen. We want to keep interaction patterns consistent. Why? Whenever we introduce an interface that behaves very differently, we are asking users to rebuild their mental model, and that takes effort. We burn a lot of their brain calories, and that is our cognitive load. So unless you have a very good reason, don't make our brain work harder than it needs to.

Building trust through transparency

[00:13:40] The second principle is building trust through transparency. Earlier we talked about how the brain reacts when something feels like a black box. In this experience, when users expand the detail we don't just show more information. We show where we found this answer and how this answer got generated. What we do is provide the source behind the recommendation so that users can see exactly where the information comes from. We also incorporate the system confidence level, which tells users how certain the AI is about the answer. Of course, if the number is high we tend to think, okay, this is probably a good resource that I can use. And third, you tell them the reasoning behind the recommendation.

[00:14:34] This matters because our brain is constantly evaluating uncertainty. When AI feels like a black box people hesitate, but when the reasoning becomes visible and available to people the system becomes more understandable. That's how we build trust: we start from transparency.

Keeping the human in control

[00:15:01] The third principle is keeping the human in control. A very helpful way to think about this is that AI should act as a copilot, not an autopilot. Our goal isn't replacing human intelligence, it is really augmenting it. What does this look like in practice? The user can take the initiative for important actions, but the AI can suggest, it can guide, it can offer recommendations, and it never forces a decision. It never takes control. A simple way of thinking about this is: AI suggests and humans decide.

[00:15:48] This is just as important because people still want to be in control. People want to have the ability to undo, to refine or to reject some of the AI's proposals, because you want to make sure that people don't feel trapped by the AI's decision. When that happens, that's when we lose trust. So always give people the escape hatch, always give them the control, just like a real copilot: the pilot is still flying the plane.

One entry point instead of many

[00:16:23] With everything that we experienced, we took a step back and really looked at the system as a whole, and we realized something important. The goal again is not just to add another help option. The goal was simply to simplify how we help people. Instead of asking users to choose between multiple paths, like the search you saw earlier, the documentation, the FAQ, support or even a live agent, we make the copilot the starting point. Now when someone needs help the copilot becomes the guide, and it can answer simple questions directly. Of course, when people need more help we route the user to the right place.

[00:17:08] So instead of navigating the help ecosystem on their own, we create a system that helps them find the right path and get their task done faster. This matters because our brain is consistently filtering what is important to us. By creating one entry point we reduce decision friction and make sure that getting help is simplified and more intuitive.

Designing for how people think

[00:17:42] If there's one idea I would like to leave you with today, it is that designing AI experiences is really not just about making the system more intelligent, it's about designing for how people think, because eventually people are the ones who use the system. The brain is constantly filtering all the signals for relevance and safety and control, and when we design we want to keep that in mind. When we reduce cognitive load, when we make the system transparent, and when we keep the human in control, AI will stop feeling like a black box. We can treat it as a trusted partner. In a way that brings us back to the restaurant menu I mentioned earlier in the talk: good design doesn't want to overwhelm people with more options, it helps them see the right choice clearly. The real goal of designing AI is for the human brain. Thank you.

Q&A

[00:18:56] Host: Excellent. That was a really, really interesting case study. I love the way you took us through almost an evolution of chatbots through it as well, because I know, and I'm sure a lot of other people out there are the same, when I encounter a chatbot my first instinct is to write "talk to a person".

[00:19:15] Julianna: Yes.

[00:19:17] Host: Because they've been so poor in the past.

[00:19:18] Julianna: Yes, yes. There are so many factors we need to consider, and one thing I felt is that a lot of the time we are trying to be useful, but we have to make sure that we are useful in a way that actually helps instead of confusing people. The search and all this FAQ and contact are there because we are trying to hand-hold our user, but we forgot we are living in a tech-savvy world. Sometimes you don't have the budget to do research, but to me, creating the best UX is really about honing in on how people think. You will never be wrong, because if you interview five people they probably have five different answers, but when you start reading about how the brain acts and works I think you will get a more accurate answer.

[00:20:11] Host: Makes sense. I have two completely polar opposite questions coming out of that. I'm going to start with one, because I loved your example of the website and it's so standard now. You've got your banner, you've got your call to action, you've got your chatbot, you've got all that.

[00:20:30] Julianna: Yeah.

[00:20:31] Host: When you were explaining that, the first thing that jumped into my head was that study a few years back where they were saying if you make your text hard to read, so if you intentionally add friction...

[00:20:42] Julianna: Yes.

[00:20:44] Host: It actually makes people, because as you were saying, the amygdala goes, this is something I'm not expecting. Because if it's something you're expecting you just brush right past it.

[00:20:52] Julianna: Yeah.

[00:20:53] Host: So is there a lesson there that we can learn, that maybe doing something instead of trying to make it smooth and instead of trying to make it as easy as possible? Is there a lesson in adding friction?

[00:21:06] Julianna: Yeah, next talk. I'm pretty sure, because like you say, a lot of the time when people say hey, I want to call attention very quickly, they think about surfacing instead of pulling it back into the background. And then when you have so much messaging that you want to put out there, everything is important. If you have different stakeholders, everything is important. But now, what is the most important thing? You don't want to say, hey, I know this button is important, and then when you go back to other people you say, I know this button is important. How do you handle that? So the best way for me is usually to pull it back. I'm going to pull you guys back and give the space for people to breathe and to digest. That's why it's very famous when people say if you cannot think about anything, go have a walk. That's when you give space. You know what I mean?

[00:22:07] Host: Yeah. And sorry, I forgot to mention to everybody watching, because I can see there's quite a few people online watching: please write in whatever questions you have and Julianna will be happy to answer them. But I'm going to jump to my opposite question now. I love the confidence rating, because that gives... the biggest fear is always the whole, has this just hallucinated or is this actually real? But just to your point of people still wanting control...

[00:22:43] Julianna: Yeah.

[00:22:43] Host: ...you were saying that okay, we're going to give them the information but they're still the driver. They're still the one who's taking the action.

[00:22:51] Julianna: I have noticed.

[00:22:53] Host: And I'm going to take the analogy of coding in the past three months.

[00:22:58] Julianna: It has shifted.

[00:23:00] Host: I would say in the programming space, where people don't want to do any of the coding. And Claude Code has this dual mode where it has planning and then it has actions. So you get the plan, it tells you this is what I'm going to do.

[00:23:16] Julianna: Yeah.

[00:23:17] Host: And people review it and go, yeah, happy with that, and then it goes and does it. Do you see that similar kind of evolution happening in maybe the chatbot space?

[00:23:30] Julianna: Chatbot's not the right word, but like a helper on the website or on the application. Yeah, I'm sure that's going to... For right now, what I'm talking about today is what I'm seeing more like in the next two years, because AI is not as robust as we think yet. But in the next two years everything is going to change. AI will be more robust than ever. And by the time... our brain is also very lazy sometimes, we have a lazy brain, it's just human beings. But by the time AI becomes so robust, are we going to let AI make decisions for us? This is a question I always ask. Because right now we let AI do competitor studies. We let AI do analysis. And what happens down the road? Go back to your coding, using AI for the code. One thing that I would always ask developers is, yes, you get all this code from AI, but is that code more lines than what you can do? When you add more lines to the code you affect the performance on your site, but is that what you really want? So a lot of the time I would encourage developers to go back and double check. It is very important that we double check, to make sure that AI is spitting out accurate answers. You know what I mean. Sometimes you think using AI simplifies a lot of stuff, but it's not, and if you really pay attention you're just like, oh, that's not right. One other thing that I think is very useful: if you really use AI to help with your work, please use multiple tools, multiple channels, to cross-check. Because if not, we're going to... yeah. So that's that.

[00:25:25] Host: I've definitely found that. I do OpenAI and Claude and have them sanity check each other's work, and always find improvements.

[00:25:36] Julianna: Yeah, right, yeah.

[00:25:37] Host: And do you think longer term, so I assume you're using one of the big models under the hood for your chatbot there. Do you think longer term that these agent kind of things, like as you said it's maybe two years down the road, but is that something that every company is going to program their own version of, with all of their context of their application? Or do you think the models are just going to become this open Claude helper who can do anything?

[00:26:07] Julianna: Yes, I do. I do see it. Even with our company, we do allow this. There are two versions of copilot. One is a basic standard one and the other one is a more advanced and robust one. We allow employees to create their own agent. So it's going to be... again, I would not say that it's going to take over people, because humans are mainly the boss of this.

[00:26:39] Host: Is accountable always.

[00:26:41] Julianna: It is, it is. And sometimes it drives me a little bit crazy when I hear people say that oh yeah, it's going to take over your job. No, that's not true, only if you let them do it. You don't, right? Because human intelligence, I don't think anything can replace that, and we are different. AI is just, to me, we keep feeding AI. It's not coming from them originally. It's the code that you put in there to create this agent. I do see that it's going to be really... use it wisely, broadly. I'm worried because I'm looking into all these icons that we have to design for these agents, and we have 14,000 people. Am I going to see 14,000 different AI icons? No.

[00:27:29] Host: Yeah. No, it's an interesting point. So yeah, I think that brings us actually to time. But I really enjoyed the case study. I thought it was brilliant, kind of outlining how it's actually working in practice today and solving some of those confidence issues and things like that. So yeah, thank you very much.

[00:27:45] Julianna: Thank you for having me.