Content Benefit

Content Strategy | UX Writing | Content Design | Globalization | Communications

Generative UI: always adapting

Generative UI: always adapting

For years, digital design has worked on a simple premise: someone designs a screen, tests it, approves it, and that screen stays identical for thousands or millions of different people. A catalogue of fixed states, planned in advance to cover every imaginable use case. Generative UI turns this model upside down: the surface people interact with doesn’t exist until it’s needed, and gets assembled on the spot by an AI model based on who’s using it, what they’re trying to achieve, and in what context.

This isn’t just heavy personalization, the kind already familiar from product recommendations or social media feeds. Here the whole production logic changes: there’s no longer a single master screen to adapt with variables, but components (buttons, forms, charts, maps) assembled in real time from an intent expressed in natural language.

What changes compared to traditional design: from fixed screens to real-time composition

The starting point of so many recent design debates, like the one around Trenitalia’s redesign that wasn’t needed, stays the same: who decides how a visual element changes, and for which audience. Generative UI shifts the question: what if the system itself made that call, case by case, guided by rules set in advance?

Nielsen Norman Group, one of the most cited references in user experience research, put forward a definition back in March 2024 that remains among the most widely used in the field: generative UI is an interface dynamically generated by an AI system to deliver an experience built around the needs and context of the person using it. Kate Moran and Sarah Gibbons, the article’s authors, offer an example that captures the scale of the shift: a flight-booking app that adjusts its presentation for a user with dyslexia, who prefers window seats, avoids red-eye flights, and watches the price closely. Instead of a single screen optimized for the average traveler, the system composes as many variants as there are combinations of real needs.

The two authors make a point worth repeating: the role of designers changes radically. Instead of drawing specific components one by one, teams set rules, constraints, and goals, so-called guardrails, within which the algorithm can operate with a degree of autonomy. Design work looks less and less like sketching a screen and more like writing a script that AI interprets differently every time.

What academic research says about AI-generated interfaces

A workshop paper presented at CHI 2026, the leading global conference on human-computer interaction, co-authored by Siân Lindley and Haijun Xia among others, defines generative UI simply as “interfaces created by AI models” and argues that fields like HCI and design need to rethink methods and workflows in light of this possibility. There’s no settled empirical evidence yet, but the academic community is already trying to steer this shift rather than just react to it.

A more concrete contribution comes from research by Bektur Ryskeldiev and colleagues, carried out between Mercari R4D and the University of Tsukuba: six empirical studies run between 2022 and 2025 with blind, low-vision, and older users of peer-to-peer marketplaces. The findings are stark. When product pages, often cluttered and written by sellers with no technical skills, get rebuilt on the fly by a model (GPT-4o in this case), scores from screen reader users jump from 3.14 to 5 out of 5, and the time needed to complete a task drops from roughly 130 to 25 seconds. An AI-based photography aid helped blind sellers frame their product photos more accurately, cutting centering error from 127 to 46 pixels. The gain goes well beyond aesthetics and touches on participation that was previously out of reach: according to the same study, 70% of visually impaired people avoid selling online because of the practical barriers the interface puts in their way.

Who’s already building it: from dev kits to apps inside ChatGPT

Vercel, the company behind the Next.js framework, added generative UI to its own AI SDK, letting developers have a language model generate not plain text but actual React components, rendered server-side and streamed to the browser as they’re produced. The topic overlaps with vibe coding, where AI already writes code from loose, informal instructions. The most cited example is still the weather one: someone asking what the weather will be like tomorrow in Milan doesn’t get a sentence back, but a small widget with temperature, icon, and forecast, generated on the spot by the model that called the right tool.

OpenAI took a similar path with the Apps SDK announced in October 2025, which lets third-party services show up inside ChatGPT with interactive interfaces, maps, playlists, and presentations, instead of plain text answers alone. Zillow, the US real estate platform, is among the first partners: Josh Weisberg, the company’s Head of AI, said the Zillow app in ChatGPT “shows the power of AI to make real estate feel more human,” calling it “a conversational guide that makes finding a home faster, easier, and more intuitive” for millions of people. A site to be navigated filter by filter gives way to a surface that emerges directly from the conversation. The same shift, from static optimization to an answer assembled on demand, already showed up in the move from SEO to GEO, where content no longer just gets indexed, but summarized and recomposed on request by an answer engine.

What companies testing it in the field are saying

Lisa Murkin, senior product manager at consultancy Elsewhen, describes two concrete cases she’s worked on in an article for Mind the Product: a large retailer’s product page, which shows large images and styling suggestions for a dress but switches to shade comparisons and virtual try-on for a lipstick, and an insurer’s claims platform, where a car damage claim triggers a lean form while a life insurance claim opens extra fields on medical history and beneficiaries. Murkin is also quick to flag the flip side: hyper-personalization can create inconsistencies across different people’s experiences, complicating the job of marketing and brand teams used to a single, consistent message for everyone.

The risks of a surface that keeps changing

The same problems that affect any system built on language models remain unresolved: hallucinations, biases inherited from training data, privacy concerns tied to a level of personalization that assumes a lot of knowledge about the person browsing. Nielsen Norman Group adds a less technical but equally concrete observation: visual consistency is one of the foundations of usability, and a surface that shifts every time risks confusing people used to finding buttons and menus in the same spot. There’s also a non-trivial computing cost: assembling elements on demand, instead of loading pages that are already built, requires processing power that, according to the same authors, could slow large-scale adoption for years. These limits sit alongside the ones already covered when discussing struggles with using AI in other corners of digital work.

Toward a design that orchestrates instead of drawing

What emerges from research, shipped products, and early corporate testimonials is a field still testing the boundaries of this technology, more than mastering it. The examples with the clearest measurable impact, like the study on screen reader users in peer-to-peer commerce, are for now concentrated in specific niches where the benefit clearly outweighs the risk of a less predictable experience. Major platforms, from Vercel to OpenAI, are meanwhile building the technical infrastructure that will make this approach available at a much larger scale. Anyone designing digital products would do well to watch this evolution closely, because their role, rather than disappearing, seems set to move from drawing the single element to defining the rules that element, generated by a model, will have to follow.

My own take, after going through the research and the corporate cases: personalization and targeting have long decided whether a digital experience succeeds, and generative UI pushes that logic to its extreme. The real challenge will be holding onto a baseline structural consistency and identifying a handful of sensible variables that actually determine what each user sees, without letting the possible combinations multiply endlessly. Will the industry pull it off? Future generations will be the judge.

Related sources

Chiara Bonifazi