In March 2024, Sarah Gibbons and I published one of the first formal definitions of generative UI (genUI). We distinguished genUI (real-time, AI-generated interfaces customized to individual users) from AI-assisted design (using AI to help designers build traditional interfaces faster). The key distinction is when and who benefits: AI-assisted design helps during development, but the end user encounters a traditional interface; with genUI, AI generates interface elements in real time and the end user's experience is what's being shaped.

We kept timelines vague because we didn't know when these interfaces would become reality. Two years later, true genUI is no longer purely theoretical. While there are a myriad of examples of experimental genUI right now, the most notable progress has happened within AI chat interfaces. Simple interactive elements — buttons, form fields, checkboxes — are starting to appear contextually within conversations, generated by the AI when it determines they'd be useful.

In This Article:

Checkboxes in Google AI Mode

Consider a traveler using Google AI Mode to research hotel options for a trip to London.  Google AI Mode returned several recommendations, each neatly formatted with thumbnail images of the hotels and — most exciting — checkboxes next to each one.

A screenshot of Google AI Mode on mobile showing London hotel results with selectable checkboxes. Under "Luxury & Landmark Hotels," three hotels are listed with thumbnail photos, prices, ratings, and brief descriptions. Blue checkboxes next to The Ritz London and The Mayfair Townhouse are selected. At the bottom, the two selected hotels appear as chips above the chat input field.
Google AI Mode added checkboxes to hotel results. When the user selected hotels, those choices appeared as chips in a bar above the chat input — ready to reference in followup questions without retyping.

Why am I so amped about a humble form-field element? Because checkboxes are useful here in the same way they’ve always been useful in classic graphic user interfaces — they allow the user to directly select specific options, rather than typing them out.

In the Google AI Mode example, selected hotels were populated into the chat input, ready for followup questions. The user didn't have to read, memorize, and retype hotel names to continue the conversation. As ever in design, this shows how a simple reduction in friction represents a major improvement to the overall experience.

Buttons in Claude

In an even more complex example, Claude sometimes generates a set of interactive form fields to collect additional context from the user before generating a response. This is a predesigned module called AskUserQuestion that Claude can decide to load and populate when it needs the user to answer more questions before it proceeds.

In the screenshot below, Claude asks the user about their budget for the trip to London. It’s question one of four — and the AskUserQuestion model seems to currently limit its use to four questions max. This is an excellent example of providing tightly controlled guard rails around an AI-generated design, which we believe is necessary to ensure a consistent and sane experience right now. Otherwise, a model may (for example) unexpectedly decide to request a user answer 35 questions in one sitting.

A screenshot of Claude responding to a hotel recommendation request by generating a multiple-choice question widget. Claude displays question 2 of 4: "What's your budget per night for this trip?" with three selectable options ($250–400/night, $400–600/night, $600+/night) and a "Something else" write-in option. A Skip button appears at the bottom.
Claude generated a multiple-choice clarification widget to understand the user's hotel preferences before making recommendations. The questions were tailored to gaps in the system's knowledge about this specific user's trip.

The system can decide to employ a variety of question formats, with each question on its own page (like any quiz, survey, or form wizard). In the screenshot above, it presents single-select multiple-choice options to ask about the traveler’s budget. It also includes an open text field for the user to provide additional context or a different response.

Context Is Critical

Like the checkboxes in Google AI Mode, Claude’s AskUserQuestion is pulling together very simple design elements to conveniently collect needed details and move the conversation further. This has become a critical component of modern AI chat experiences.

It was never realistic to expect consumers to become perfect prompt engineers, comprehensively planning out the necessary context they’ll have to provide upfront to achieve their goals. We’ve trained people for years to use short keyword phrases to navigate the digital world, and it shows in how people are prompting. In our studies, most people (other than early adopters and AI enthusiasts) are using very simple prompts like, “Help me plan a trip to London.” Without more information (the user’s price range, interests, available dates, length of trip, etc.) the system can deliver some helpful overview information. But that approach leaves the power of AI chat (customized responses) untapped.

So, these followup questions are critical, but without interactive genUI elements, they’re very tedious to provide. Consider Perplexity’s response to a vague prompt in the screenshot below. It asks the user to provide detailed answers to 5 important questions, so it can give a better response.

A screenshot of a Perplexity AI chat conversation. The user prompts, “Help me plan a trip to London”. The system responds by asking for more details to help it improve its response. The questions are presented in a numbered text list. For example: “1. When are you thinking of going, and for how many nights? 2. Approx budget per night for lodging (USD or GBP)?”
In a response to a vague prompt (Help me plan a trip to London), Perplexity asked a series of 5 followup questions, presented as text. (It prepopulated some of the suggestions based on what it already knows about me. For example, it knows I live in central North Carolina, so RDU is my closest airport. While not an example of genUI, it is an example of using AI’s memory to improve the quality and specificity of its output.

To respond, users must type out their answers to each question one at a time, often relying on their short-term memory. If they didn’t recall the questions, they’d need to shift their attention back and forth between the prompt text box and the set of questions, making sure to write the correct number for each question. The process is error-prone and can become cognitively taxing over multiple interactions, causing users to skip important questions or give incomplete answers.

Compare that experience to answering the same set of questions through Claude’s AskUserQuestion widget. With those simple form fields and multiselect buttons, the experience of providing these details becomes substantially faster and easier. The user is walked through each question, one by one. For simple questions with a limited set of responses like #3 (Who’s traveling?), just selecting an option is much simpler. And for more complex questions, the user should be able to type their answer in an open-text field immediately below the question.

AI Chat Is Where People Interact with AI (Right Now)

These improvements may seem minor. Checkboxes, buttons, and form fields are some of the oldest UI elements. But, when deployed in the right place and at the right time, they reduce unnecessary friction in the user experience. In AI chat, they prevent the back and forth that makes chat feel slow or tiring.

This matters because chat is where most people interact with genAI — so even small interaction upgrades can improve outcomes for a massive audience, beyond the power users or AI enthusiasts experimenting with more exciting, cutting-edge genUI.

These are simple design patterns that people already understand and use. They help the system gather the context it needs without forcing users to plan ahead, remember a list of questions, or format their answers perfectly. As genUI evolves, its usability will rely on such interactions that help users clarify intent and that improve AI outputs over time.