Generative UI (research.google)

Google Research announced "Generative UI," a new capability—now rolling out as experiments in the Gemini app (dynamic view and visual layout) and in Google Search’s AI Mode—that lets Gemini 3 Pro build complete, interactive user interfaces on the fly in response to any prompt. Rather than returning static text or images, the model generates bespoke visual experiences (web pages, games, simulations, tools) and even the code to run them, tailored to intent (e.g., teaching a child vs. briefing an expert). The feature leverages Gemini’s multimodal understanding and agentic coding to create interfaces optimized for learning, exploration and task completion, and is available today to Google AI Pro and Ultra subscribers in the U.S. via the “Thinking” model in AI Mode. Technically, the implementation augments Gemini 3 Pro with three key components: controlled tool access (image generation, web search, etc.) served by a backend, carefully crafted system instructions (goals, plans, tool manuals and formatting rules), and post-processing to mitigate common errors. Google evaluated Generative UI against human-designed sites, top search results, and raw LLM outputs using a new PAGEN dataset; human experts still lead, but generative UI was strongly preferred over standard LLM outputs (speed excluded). Limitations include slower generation (often minutes) and occasional inaccuracies; future work will focus on speed, broader service integration, style consistency and interactive adaptation from user feedback.