Journal · Field Notes

The interface draws itself.

For most of software's history the interface was a fixed artifact. A designer drew every screen, an engineer built it, and the product shipped with its surface settled. That assumption is now breaking. Generative UI has moved from demonstration into production: with tooling like the Vercel AI SDK, the model composes the interface at runtime, assembling it from a person's intent instead of rendering a layout that was frozen months earlier. The screen is becoming an output of the system rather than a thing drawn once and kept. This sounds like a story about designers losing work. It is the reverse. When the screen can be generated, the value moves to the thing that governs the generation, and that thing has to be designed more deliberately than any single screen ever was.

2026.067 minIssue 12

Begin with what generative UI actually changes. In the old model, a layout was authored. Every state, every button, every empty screen was placed by a person and committed to the codebase. In the new model, the person authors a system and a set of intentions, and the model draws the specific surface a given user needs in the moment. The components still exist, but which ones appear, in what arrangement, for whom, is decided at runtime. The interface stops being a noun the team owns and becomes a verb the model performs. That shift is small in code and large in consequence, because it relocates where the design work lives.

The relocation runs straight into the year's loudest design finding. Nielsen Norman Group's State of UX 2026 reports that AI adoption is climbing while trust is falling, and states the strategic move plainly: the competitive moat is shifting from model performance to user experience, because the model underneath is becoming a commodity. A person who was burned by one AI feature approaches the next one with their guard up. So a generated interface inherits a trust problem the moment it appears. It can render anything, which is exactly why the person on the other side needs to believe it will not render the wrong thing. Capability without trust does not convert. It stalls.

This is why the system, not the model, is the asset. A rented model can draw a screen. It cannot, by itself, decide what this particular person should be allowed to see, change, or undo, or when it should stop and hand control back. Those decisions are the design system in the fullest sense: the components the model may compose, the constraints it must defer to, the states it is forbidden to invent, the boundary where automation yields to a human. None of that is generated. All of it is drawn, once, with care, by people who understand the business the interface serves. The model improvises inside the system. The system is the score.

A rented model can draw the screen. It cannot decide what a person should be trusted to see, change, or undo. That decision is the drawing, and the drawing is the moat.

Notice what becomes load-bearing under this arrangement. The most valuable artifact is no longer the highest-fidelity mockup of a single flow. It is the set of rules and primitives the generation lives within: a component library with real semantics, clear affordances for control and reversal, honest signals about what the system did and why, and graceful behavior when it fails. These are the fundamentals trust is built from, and they are now the difference between a generated interface a person leans on and one they quietly stop using. A studio that has drawn that system well can let the model compose freely. A team that has not will find the model exposing every gap they left.

The practical reframe for a business leader is to stop evaluating AI features by the cleverness of the model and start evaluating them by the quality of the system you would let that model draw inside. The studio has made adjacent versions of this argument before. The design system is the substrate that decides whether anything built on top of it holds, and the model itself is a rented part you do not own and cannot differentiate on.

Put those together and the conclusion for the generative-UI era is direct. If the model is rented and the screen is generated, then the one durable asset is the system you draw for both to live inside, and trust is the property that asset has to earn. That drawing is yours. It is the thing a competitor cannot rent off the same price list, and it is what turns a model that can render anything into a product a person is willing to rely on.

Field note prompted by the move of generative UI from demonstration into production, where tooling like the Vercel AI SDK composes interfaces at runtime, and by Nielsen Norman Group's State of UX 2026, which reports AI adoption rising while trust falls and names the shift of the competitive moat from model performance to user experience. The principle outlasts the tooling: rent the model, draw the system, earn the trust.