Is chat the wrong default interface for AI products?

Medeiros argues that "the wholesale adoption of conversational interfaces as the default for AI products is not an innovation — it is a regression" that "risks discarding four decades of hard-won progress in human-computer interaction" and "erects new cognitive barriers precisely where knowledge workers need the least friction." The piece opens on a product demo where every action — adjusting a budget, assigning a task, rescheduling a meeting — must be typed rather than clicked or dragged: "The interface has become more expressive but less efficient. It is more capable, in a narrow sense, but harder to use." Medeiros grounds the claim in "three bodies of research: Shneiderman's theory of direct manipulation, Norman's action theory and the gulfs of execution and evaluation, and Sweller's cognitive load theory," and cites one named empirical comparison: "Valtolina et al. (2019) compared GUI-based interfaces with chatbot interfaces across structured, goal-directed tasks. GUI-based systems produced consistently higher communicability and task performance."

The mechanism Medeiros gives is that conversational interfaces reintroduce the exact problem direct manipulation was invented to remove. Shneiderman's 1983 paradigm rested on "continuous visual representation of objects of interest, physical actions in place of complex command syntax, and rapid, incremental, reversible operations whose effects are immediately visible" — properties that let users "concentrate on their tasks rather than on the medium." Chat restores a symbolic layer between intent and action: a user must "compose a request, verify it was understood, and reconstruct from a text response whether the right thing actually happened." From this, Medeiros proposes treating modality as "a deliberate design decision rather than an implementation convenience," gated by two questions: "could this be done faster by clicking than by describing, and does the value lie in the outcome or in the visible human process that produces it?" The second question protects conversation where it is actually warranted — tasks whose value lies in a visible human process, not merely an outcome — rather than banning chat outright; the prescribed remedy is "hybrid patterns" that mix conversational and direct-manipulation elements depending on which failure mode a task poses.

This both parallels and pushes against How should AI interfaces handle the shift from doing to supervising?: Nielsen keeps a direct-manipulation layer only as a fallback reserved for edge-case editing, granular corrections, and emergency overrides, while Medeiros's gate questions would promote clicking to a first-class default for any task that fails the value lies in the outcome test, not just for correcting agent errors. It runs on the same cognitive-load lineage as Is the AI capability gap really an interface problem?, which cites a different cognitive-load study to make the same not-model-capability argument — but Mollick frames the fix as new interfaces (specialized, agentic, generative) that will arrive and feel like capability leaps, whereas Medeiros frames it as a design-discipline failure to apply theory the field already has. The empirical finding behind Do generated interfaces outperform text-based chat for most tasks? is the kind of result Medeiros's theoretical case predicts, though it measures generated UI against static chat rather than click against chat specifically. And Do generated analysis UIs really work better than chat? supplies a finer-grained, sub-task version of the same trade-off Medeiros argues at the whole-product level: GUI structure buys clarity at the cost of rigidity, so neither modality is categorically superior — the gate questions are one way to decide, task by task, which cost to accept.

The excerpt is a design essay published on a personal/professional blog, not a peer-reviewed study, and it does not report new data of its own; the only study named, Valtolina et al. (2019), is summarized in one sentence without sample size, task count, or effect size given here. What the excerpt supports, at the strength of a practitioner argument built on established HCI theory plus one cited comparison, is that chat-as-universal-default lacks the theoretical backing often assumed for it — not that any specific hybrid pattern or gate-question application has been shown to outperform chat-only design in practice.