Generative UI and the Ephemeral Interface

This week, Google debuted their Gemini 3 AI model to great fanfare and reviews. Specs-wise, it tops the benchmarks. This horserace has seen Google, Anthropic, and OpenAI trade leads each time a new model is released, so I’m not really surprised there. The interesting bit for us designers isn’t the model itself, but the upgraded Gemini app that can create user interfaces on the fly. Say hello to generative UI.

I will admit that I’ve been skeptical of the notion of generative user interfaces. I was imagining an app for work, like a design app, that would rearrange itself depending on the task at hand. In other words, it’s dynamic and contextual. Adobe has tried a proto-version of this with the contextual task bar. Theoretically, it surfaces up the most pertinent three or four actions based on your current task. But I find that it just gets in the way.

When Interfaces Keep Moving

Others have been less skeptical. More than 18 months ago, NN/g published an article speculating about genUI and how it might manifest in the future. They define it as:

A generative UI (genUI) is a user interface that is dynamically generated in real time by artificial intelligence to provide an experience customized to fit the user’s needs and context. So it’s a custom UI for that user at that point in time. Similar to how LLMs answer your question: tailored for you and specific to when that you asked the original question.

Near the end of the short article, they point out some challenges, including usability.

Constantly changing UIs will cause usability problems. Much of users’ understanding of modern web interfaces is rooted in design standards (for example, logos are often in the top left). The more you use a website, the more familiar (and thus efficient) you become. As Gen UI alters the interface based on your needs, you could be shown a different UI every time you use a website. This constant relearning of the interface might cause frustration, especially in the beginning, as users transition from the old ways.

That’s been my concern all along. In fact, consistency is number four in Jakob Nielsen’s Usability Heuristics.

From NN/g: GenUI offers the potential to shift from single-experience design to personalized experiences for each individual.

In the same genUI article, Kate Moran and Sarah Gibbons share a speculative example of a user booking a flight. The system knows that the user “never takes red-eye flights, so those are collapsed and placed at the very bottom of the list.” And that she prioritizes cost and travel time so those datapoints are displayed more prominently and the results are sorted accordingly. Can you imagine the support nightmare for UI that’s always changing? How would a support person even begin to troubleshoot?

But let’s get back to this week and Gemini.

Content Is King

We’ve seen glimpses of genUI in both ChatGPT and Claude before. Instead of on-the-fly UI, we’ve been getting on-the-fly tool calling. In other words, for some queries, ChatGPT might open up the canvas, write some code, return a chart, or generate an image. The content determined its display. But, of course, 95% of it was text and emojis.

With the new Gemini app, Google takes this idea further by creating interactive content or experiences in realtime to answer queries. In their blog post about generative UI, Google uses the example prompt of “Create a Van Gogh gallery with life context for each piece.” The result is a visually-rich microsite.

Example of generative UI in Gemini based on the prompt, “Create a Van Gogh gallery with life context for each piece”

In another example, an interactive widget appears at the top of a Google search. It’s a colorful illustration of how RNA works. Gemini is doing the work of a designer here—figuring out the best visual and interactive way to communicate an idea.

In other words, the generative UI in Gemini isn’t the chrome or frame around the experience, it’s the content. It’s almost no different than a TikTok feed where it could be a regular video, a live shopping stream, or and ad.

Visually Trending

By having the AI “design” behind the scenes, it begs the question, how does it do on taste? From my professional standpoint, I think it’s a solid B or B+ letter grade. In their research paper about generative interfaces, the researchers attempted to quantify that taste by having study participants rate their preferences against human-crafted websites, formatted text, the top search result webpage, or plain text.

From the blog post:

The sites designed by human experts had the highest preference rates. These were followed closely by the results from our generative UI implementation, with a substantial gap from all other output methods.

Gallery of generative UI examples from Gemini

In my clickthrough of the generated websites, I can see some common UI patterns and components. There’s the rounded rectangle with a thick border on just one side. Oh hey, it’s Playfair Display paired with Lato again. And pill-shaped buttons everywhere! Essentially, the resulting generated interfaces look like the work of about three or four mid designers but lack the sophistication of seasoned pros.

But you know what? That’s OK for this use case. These interactive experiences are completely ephemeral.

I would recommend that the visual styles get updated every year or so, lest they become stale like the default templates in Google Slides.

(I think it’s also worth noting that the “human experts” are highly-rated freelancers from Upwork who were paid $100–130 per smallish website which took them, on average, three to five hours to complete. This is disclosed on page 15 in the paper. Take that how you will.)

How to Train an AI Designer

At the bottom of page 15 of the research paper is the start of the system prompt. It’s a fascinating four-and-a-half page read.

The first bullet under “Core Philosophy” demands an interactive-first approach:

**Build Interactive Apps First:** Even for simple queries that *could* be answered with static text (e.g., "What’s the time in Tel Aviv?", "What’s the weather?"), **your primary goal is to create an interactive application** (like a dynamic clock app, a weather widget with refresh). **Do not just return static text results from a search.**

It later goes into the how by specifying Tailwind CSS, HTML Canvas, and SVG.

There’s an “internal” thought process that somewhat mirrors the human creative process:

1. **Interpret Query:** Analyze prompt & history. Is search mandatory? What **interactive application** fits?
2. **Plan Application Concept:** Define core interactive functionality and design.
3. **Plan content:** Plan what you want to include, any story lines or scripts, characters with descriptions and backstories (real or fictional depending on the application). Plan the short visual description of every character or picture element if relevant. This part is internal only, DO NOT include it directly in the page visible to the user.
4. **Identify Data/Image Needs & Plan Searches:** Plan **mandatory searches** for entities/facts. Identify images needed and determine if they should be generated or searched, as well as the appropriate search/prompt terms for their ‘src‘ attributes (format: ‘/image?query=<QUERY TERMS>‘ or ‘/gen?prompt=\<QUERY TERMS\>‘).
5. **Perform Searches (Internal):** Use Google Search diligently for facts. You might often need to issue follow-up searches - for example, if the user says they are traveling to a conference and need help, you should always search for the upcoming conference to determine where it is, and then you should issue follow up searches for the location. Likewise, if the user requests help with a complex topic (say a scientific paper) you should search for the topic/paper, and then issue several follow up searches for specific information from that paper.
6. **Brainstorm Features:** Generate list (~12) of UI components, **interactive features**, data displays, planning image ‘src‘ URLs using the ‘/image?query=‘ format.
7. **Filter & Integrate Features:** Review features. Discard weak/unverified ideas. **Integrate ALL remaining good, interactive, fact-checked features**.  

But most surprisingly, there’s only a single paragraph directing the quality and design:

**Sophisticated Design:** Use Tailwind CSS effectively to create modern, visually appealing interfaces. Consider layout, typography (e.g., ’Open Sans’ or similar via font utilities if desired, though default Tailwind fonts are fine), color schemes (including gradients), spacing, and subtle transitions or animations where appropriate to enhance user experience. Aim for a polished, professional look and feel. Make sure the different elements on the page are consistent (e.g. all have images of the same size).

Not Cooked Yet

Google Gemini’s implementation of generative UI might be 2025’s answer to HyperCard that I’ve been asking for. Although all these UIs are ephemeral rather than lasting like HyperCard stacks. The important thing, however, is that main interfaces aren’t being dynamically generated and wreaking havoc on usability. Generating content UI, or really _experiences_seems to be the right application of this technology. I find it akin to the interactive charts that The New York Times sometimes produces for their news stories.

But with AI making these decisions about how to present content compellingly in a visual and interactive way, I suppose the question on every designer’s mind is “Are we cooked yet?”

In his Substack post about the news, Jakob Nielsen seems to think that AI-generated UI “will be better than human-created UI design by late 2026,” at least for the simple stuff like we’re seeing from Gemini 3. But he goes on to extrapolate that this capability will double every seven months or so, therefore “we should expect AI to get about 4 times better each year in doing UX design and user research.”

I don’t think I buy that. Maybe he’s right for individual deliverables like wireframes, task flows, or synthesizing research.

Regardless of how good AI might become, we will still need to be the puppet masters pulling the strings. We’ll still need to get alignment among humans in our respective organizations, perform systems thinking, and dispatch and monitor the AI designer agents. Our work might indeed become prompt, generate, deploy.