Does visible design structure keep HCI expertise necessary as AI advances?

In the discussion section of their just-in-time objectives paper, Lam and colleagues argue against what they call "dominant AI narratives" that "tend to automate away the work of HCI because they envision a model that absorbs and obscures its inner workings, including interface design decisions." Against that narrative, they contend that their just-in-time objectives architecture does the opposite: it "makes visible the numerous design decisions that go between an AI system and a user interface." Because the resulting design space is large and, in their words, "lacking a singular 'right answer,'" they conclude that "HCI and design expertise is critical to navigate these decisions."

The reasoning rests on where decisions land once an objective has been inferred. Inferring a just-in-time objective does not finish the job of producing a tool — Poppins, their own instantiation, still has to "generate and evaluate candidate design specifications" before it can assemble a working interface. That generation-and-evaluation step is a design choice point that a single-objective, opaque chat interface either hides or never surfaces at all. Because this choice point recurs at every interaction rather than being settled once at training time, the authors position HCI judgment as something the architecture needs continuously, not just once during model-building — a structural argument for keeping designers in the loop as AI systems take on more of the generation work themselves.

This claim runs parallel to How should AI interfaces handle the shift from doing to supervising?, which likewise argues that mature AI interaction needs explicit, designed layers rather than a single opaque model boundary — both treat visible structure, not further automation, as the route to usable AI interfaces. It also echoes the trade-off framing in Do generated analysis UIs really work better than chat?, where generated-UI choices likewise have to be actively navigated rather than resolved automatically. Read alongside the sibling note Can systems infer what users want from watching their screen?, this is the authors' own interpretation of why their measured win rates matter beyond the specific tasks tested: the architecture reframes what AI progress implies for design work rather than narrowing it.

The excerpt presents this as an argued position in the paper's discussion and conclusion, not as something the reported experiments measured directly. It does not quantify how much HCI judgment went into Poppins' own design-specification step, nor does it compare outcomes produced with and without that expertise. The claim therefore stands as the authors' interpretation of their findings' broader stakes, not as a separate empirical result — a reader should weigh it as motivation and framing rather than as evidence on its own.