Navigating the Generative UI Hype

You may recall that just last week I suggested that the UX community ignore Jakob Nielsen (and Nielsen Norman Group more generally), but this week I caught a blog post written by Nielsen Norman Group VPs Kate Moran and Sarah Gibbons that I think deserves some analysis.

But first, if you’re not a subscriber then you should be! I won’t spam you, and members can post comments and receive additional, in-depth stories a few times per month.

There’s a lot wrong with the post, and Kai Wong does a great job of covering the accessibility aspects. For this article I’m going to focus on some other perspectives, notably privacy and feasibility.

Background

The post’s central proposal is that future AI systems will be able to automatically generate hyper-personalized experiences that perfectly suit every user’s needs.

For the purposes of the article, Moran and Gibbons define generative UI (genUI) as follows:

“GenUI refers to a user interface dynamically generated by artificial intelligence in real time to provide a personalized experience.” The result of this is, “Every end user interacts with an interface built just for them and their needs in that moment.”

In other words, every time you open a website, app, or the like, you will get a completely different user interface.

They rely on a rather lengthy example, which I asked my new BFF Copilot to summarize:

Alex, a frequent flyer with Delta Airlines, uses her personalized Delta app to book a flight to Chicago for a client visit. The app is tailored to her needs, using a special font and color contrast to accommodate her dyslexia. Alex speaks aloud to the Delta AI agent to request flights from Miami to Chicago from May 6 to 10. The system searches for flight options while also checking for any weather or events that could impact her trip, alerting her to a major event that could make her travel more expensive and advising her to book early.

The presentation of Alex’s flight options is determined by her past behavior and preferences, with cost and travel time displayed prominently and results ranked accordingly. The first option would fit her needs best, but there are no window seats left, so she moves on to the next option. Red-eye flights, which Alex never takes, are collapsed and placed at the bottom of the list. Delta’s use of genUI makes it feasible to deliver such a personalized experience to each of its 190 million yearly flyers.

Intent

Any interaction relies on intent: we need to know what the user is trying to do in order to help them complete the task. Perhaps they want to watch a video, or post a comment, or book a flight. Without intent, we’re just guessing.

In the example, Alex asks the app to book a flight between certain dates. Fair enough, we have intent. But from there the AI system infers a lot. The system hides flights that it thinks Alex won’t want, and rearranges things based on what she has done in the past and what it thinks she wants right now. (Remember that AI systems are really just complicated prediction machines.)

But what happens if Alex wants something else? She never takes red-eye flights but what if there is a family emergency and she needs to get to Chicago tonight? The system has just made an already-stressful situation much more difficult because Alex has to scroll to the bottom and try to find a flight that works for her.

How would the system handle other scenarios in which the user wants to review or choose different options? Would it generate a different individualized UI? What options does it present, and what if they don’t like those options? And most important, how confused is the user when they get something that looks and behaves differently? Can they escape the AI-generated interface or do they have to struggle to figure out what prompt will get them what they want?

Privacy? What Privacy?

In this example, the system is seamlessly adapting to Alex’s needs with little or no further input. The system has to know a lot about a user in order to function in this way, and this means that the system needs a mountain of data about every single user.

This presents a massive privacy risk, and you may recall from my Ethical Design article that the vast majority of Americans are concerned about how their online data is used. Worse, it’s unclear how airlines use the data they already have. Just this week, the US Department of Transportation (DOT) opened an investigation into how airlines use the data they collect on travelers.

It’s true that providing accessibility adaptations is kind of a neat idea. But again, asking people for detailed information about personal health issues is incredibly problematic. Are you willing to trust an airline to safeguard your health information?

Data privacy and security are very real problems and users are rightly worried about them. Any product that relies on extraordinary amounts of very personal data must account for these realities in concrete, meaningful ways.

Bias, Misuse, and Heat

Moran and Gibbons do touch on a few points of concern, though only lightly. Privacy is mentioned in passing and they note that “Generative AI’s problems are GenUI’s problems.”

We know that AI systems are prone to bias and hallucinations. These are very serious issues and need to be fully addressed before deploying AI into critical situations. Imagine a user books a flight and arrives at the airport to learn that the flight doesn’t exist because of an AI error. Or it’s later revealed that AI systems regularly book people of color onto longer or more expensive flights.

They also acknowledge the computing power required for dynamic, AI-generated UIs, but only insofar as it applies to on-device computing power. There seems to be no consideration whatsoever for the environmental or climate impacts of developing and deploying massively-complex AI systems.

The hypothetical app also warns our traveler of “a major event that could make her travel more expensive and advising her to book early.” Call me a cynic (I totally am), but these systems are incredibly ripe for abuse by their owners. Yes, they could advise Alex to book early to save money, but they could also just as easily present options that are more expensive. To put it another way: what would Ticketmaster do?

What’s the Point?

After reading the blog post a few times and writing this article, I’m still not entirely sure what the point is. The authors don’t explore in meaningful depth how genUI would really work and don’t consider any technical implications. And they simply skim over some very serious problems without offering any solutions.

The fact is, some of their stated problems could be solved today with existing, non-AI technology. It’s not that difficult to use good ol’ machine learning (ML) to make predictions based on past behavior. Moran and Gibbons attempt to head this off by saying that, “This individual example may be plausible without genUI, but not at scale.” That’s hogwash. Tons of apps already offer recommendations based on past behavior. If TikTok isn’t “at scale” then I don’t know what is.

GenUI Could Work. Kind Of.

Any hyper-personalized user interface will require a flexible but consistent framework – a design system, if you will. Guidelines need to include consistent placement of certain features such as help, settings, and profiles, regardless of the overall UI configuration. Any UI needs to account for accessibility, and merely hoping that AI will figure it out isn’t sufficient. And users will need easily-accessed fallback mechanisms to accomplish tasks that differ from what the AI thinks they want to do.

In the end, this begins to look a lot like a content design problem; how do we present content in a hyper-personalized way that could improve user experiences while still presenting a consistent UI, leaving users in control, and minimizing cognitive load?

It’s certainly possible to have a system alter the UI to accommodate personal needs, such as presenting a different color set to colorblind customers. But having a system attempt to build and present a completely bespoke experience to every single user with each visit seems to bring a lot of downsides with very little benefit.

We already have the tools to create personalized user experiences; we don’t need a magical AI revolution.