Some studies say custom interfaces beat chat; others just run both together — are they actually disagreeing, or testing different things?

This explores why some studies report that visual, task-specific interfaces beat chat, while others design systems where chat and UI run together. The short answer from the corpus is that the two camps mostly measure different things. Once you see which tasks they test and which outcomes they count, much of the contradiction goes away. 'UI beats chat' usually means a particular kind of task, judged by a particular yardstick.

The strongest 'UI wins' result comes from studies of information-dense, structured work. In those studies, LLM-generated dashboards and tools were preferred over text replies in more than 70% of cases, because laying information out visually lowers the effort of reading it Do generated interfaces outperform text-based chat for most tasks?. Medeiros puts this in design-history terms: for goal-directed tasks, chat brings back the typed-command layer that point-and-click interfaces were invented to remove. Medeiros proposes two gate questions, about speed and about where the value comes from, to decide case by case whether talking or clicking fits better Is chat the wrong default interface for AI products?. That is already a conditional claim. It says 'UI wins for structured tasks,' not 'UI wins.'

The 'side by side' camp starts from a different observation: each mode has a cost the other can cover. TaskArtisan found that generated analysis widgets made results clearer but harder to change mid-task. Users swapped the open-endedness of chat for rigidity, plus the work of specifying what they wanted Do generated analysis UIs really work better than chat?. Friedman argues the reverse: chat's real bottleneck is the slow work of typing out what you want, and UI patterns like intent capture, style lenses and canvases can sit around chat to absorb that friction Can task-oriented UI patterns reduce the friction of AI chat interfaces?. Microsoft's Copilot design goes further and makes chat the coordination layer for directing several agents across apps, rather than a competitor to them Does chat-first interface actually reduce cognitive load for workers?. On the engineering side, Realtime-Venus shows that conversation and background task execution can share one ordered timeline. In that setup, 'concurrent' is literal: you keep talking while delegated work runs and its results fold back into the conversation Can frontends handle delegation while staying conversationally engaged?.

The most revealing piece is a hybrid-interface study of 73 users. Mixing AI chat into a traditional interface cut clicks, page navigations and scrolling. Task completion time didn't change across modes Does chat delegation actually save time on task completion?. Effort and speed moved independently. So a study measuring preference or cognitive load can crown UI, a study measuring physical effort can crown the hybrid, and a study measuring time can find no difference, all for the same interfaces. The verdict depends on which outcome a study chose to count.

The less obvious lesson is that the medium can change outcomes even when the AI underneath is identical. In a therapy study, a robot and a structured worksheet reduced distress while a chatbot running the same language model did not Why do robots outperform chatbots in therapy despite identical language models?. Similarly, voice interaction with ChatGPT was associated with less loneliness and dependence than text Does heavy ChatGPT use make people lonelier and more dependent?. So 'UI vs chat' is really a narrow case of a wider question: what does the form of an interaction do on its own, separate from the model behind it? The corpus doesn't yet contain a head-to-head study that tests the same tasks across both framings. That gap is why the two camps can keep talking past each other.