Can an AI-built interface feel familiar every time you return, without trapping you in a layout you can't change?

This explores whether an interface an AI generates on the fly can stay familiar from session to session without locking users into a layout they can't change. The corpus has no long-term study of generative UIs, so nothing here is measured over months. Read together, though, the notes suggest the two goals only collide when one mechanism is asked to deliver both.

The collision is real at the level of a single interface. Do generated analysis UIs really work better than chat? found that generated widgets make LLM-assisted analysis clearer but harder to modify mid-task. Easy-to-use UIs resist customization, and flexible ones ask non-programmers to think like engineers. The pull toward doing it anyway is strong: Do generated interfaces outperform text-based chat for most tasks? reports people preferring generated interactive interfaces over chat text in more than 70 percent of cases. So users want these UIs, and the rigidity cost comes attached.

Freezing the layout isn't the fix, because variation is what the medium does. Why does AI output change with every prompt and context? argues that AI output changing with prompt, sampling and audience is a defining feature, not a defect. How does AI context differ from conventional software context? adds that the context behind each screen (prompt, history, retrieved data, hidden state) keeps shifting in ways users can't memorize the way they memorize a menu structure. How should users control systems with unpredictable outputs? draws the practical conclusion. Classic consistency heuristics break here, so the goal moves from identical pixels to a user who can still predict what the system will do. Its principles (co-creation, tolerance for imperfection, support for the user's mental model) all serve that predictability.

The nearest worked example of holding something steady while everything else varies comes from a different field, persona simulation. Can training user simulators reduce persona drift in dialogue? trains for consistency directly and cuts drift by over 55 percent. It measures three separate failures: drift within a turn, drift across a conversation, and factual contradiction. That study covers dialogue simulators, not interfaces. Still, the three failures map onto a UI that shifts within a screen, across sessions, or contradicts what it did before. Can personas evolve in real time to match what users actually want? supplies the other half. A persona layer sits between memory and action and updates from recent interactions, so identity moves slowly, on evidence, instead of being regenerated from scratch each time. The stable part is a model of the user, and the surface can vary on top of it.

The generator can't be trusted to hold that stable part on its own. Why does autoregressive generation fail at constraint satisfaction? shows that autoregressive models can't retract what they've emitted, so a rule like keep the navigation where it was is a constraint they can't reliably enforce alone. Symbolic solvers help because they supply the missing check. Does AI text generation unfold through temporal reflection? points the same way: the model has no felt duration, so it doesn't remember what the interface looked like last week unless someone hands that history back. If you take these notes together, consistency has to live outside the generator, in stored user structure and explicit checks. The generator then stays free to vary the surface, and rigidity shrinks to whatever the user chooses to pin.