Can previewing AI output before submission reduce prompt iteration?

The paper argues that generative AI (GenAI) interface design has concentrated almost entirely on feedback — "helping users understand what happened after they submit a prompt" — and that this leaves users to "engage in excessive turn-taking with the AI's feedback to clarify their intent, leading to significant cognitive load and time investment." It proposes that GenAI systems also need feedforward: "designs that tell users what AI will generate before the user submits their prompt." The excerpt frames the problem as what Subramonyam et al. call the "gulf of envisioning" — the user's gap in identifying and articulating a desired task, giving effective instructions, and anticipating the model's response — and argues feedforward can "reduce excessive conversational exchanges, bridge the abstraction gap between users and AI," and "encourage meta-cognition when constructing prompts." The paper notes feedforward has an established HCI lineage outside GenAI (gestures, VR/AR, widgets, proactive systems) but says the field has only begun applying it to generative systems, mostly around assessing generated code and images.

To make the proposal concrete, the paper lists four properties effective GenAI feedforward should give users: the ability to "anticipate what the AI's response or action may be," "disambiguate their prompt without excessive exchanges," "directly engage with the feedforward content," and "reflect on their prompting practices with less friction." It then works through one instantiation in detail, a conversational LLM UI, where two feedforward components update live as the user types: a key topic phrase with a high-level outline of the anticipated response, and a "minimap view with paragraph blocks" indicating whether the response will be short or long. Editing the prompt updates both components immediately, which the paper says can surface a misalignment the user wasn't aware of — for example, realizing mid-edit that a term is ambiguous or that a response needs to be shorter — before any output is generated. The paper states it designed comparable feedforward scenarios for three further GenAI application types (document editors, malleable interfaces, and agent automations) but the excerpt does not carry those designs.

This reframes a problem the vault already has other angles on. How should users control systems with unpredictable outputs? names the same gulf-of-envisioning dynamic and treats it as something users must learn to work with — through prompt-engineering skill, mental models, and tolerance of imperfect output. Feedforward is a structurally different answer: instead of asking users to build better mental models of a stochastic system through trial and feedback, it tries to make the system's likely output legible before the user commits a prompt at all, shrinking the number of feedback cycles variability would otherwise require. It also sits one level down from How should generative UI be organized as a design space?, which organizes genUI by who it targets and how it changes over time — a feedforward preview is itself a small, ephemeral, system-generated UI, and the paper's four application domains map onto that proposal's audience and timeframe dimensions. Within the malleable-interfaces case specifically, feedforward's goal of previewing an anticipated artifact also parallels Do generated analysis UIs really work better than chat?, since both treat the interface's job as closing a gap between what a user intends and what the model will produce, one before generation and one after.

The excerpt is a short paper whose explicit purpose is to "spark discussion," not to report an evaluation: it presents designs and scenarios, not user studies, usage data, or comparisons against feedback-only baselines, and it says the three non-conversational applications (document editors, malleable interfaces, agent automations) were only described, not detailed, in this text. It does not establish that feedforward reduces turn-taking or cognitive load in practice, only that the authors judge the mechanism plausible enough to warrant "a more rigorous investigation of a design space and a set of guidelines." What follows at that strength is an open design proposal and a worked conversational-UI example, not a validated technique.