Why does typing into a chat box feel so trial-and-error — and could the interface show you the outcome before you hit submit?

This explores the gap between telling users what an action *will* do and telling them what it *did*. In classic interface design, feedback is the response that comes after you act: the button depresses, the file appears in the folder, the error message pops up. Feedforward comes before: a hover preview, a ghosted outline showing where a dragged window will snap, a label that makes the outcome predictable. With generative AI the two drift far apart. A prompt box is almost pure feedforward-free territory. You type, submit, and only then discover what the model made of your words. The collection's most direct treatment argues that this is the root of the endless back-and-forth of prompting, and that GenAI interfaces should preview anticipated output *before* submission Can previewing AI output before submission reduce prompt iteration?. It names four jobs feedforward does: anticipating the result, disambiguating what you meant, letting you engage directly with the preview, and prompting you to reflect on whether that's really what you want.

Why is feedforward so much harder with AI than with a drag-and-drop file manager? Because the thing you'd preview keeps moving. Conventional software has a fixed context you can learn, so the same click does the same thing every time. AI context, by contrast, is a shifting mix of prompt, history, retrieved data and hidden state that users can't internalize How does AI context differ from conventional software context?. If you can't predict the system from experience, the interface has to do the predicting for you. That's exactly the gap feedforward is meant to fill.

This also reframes the debate over chat interfaces. Medeiros argues that chat is a step backward for goal-directed tasks because it brings back the symbolic layer that direct manipulation was invented to remove Is chat the wrong default interface for AI products?. Seen through this lens, direct manipulation was always a feedforward technology: you see the object, you see what moving it will do. Friedman's alternatives to chat, such as node-based intent maps, style lenses and canvas editing, can be read the same way. They let you shape the output before committing, so you don't have to describe it and wait Can task-oriented UI patterns reduce the friction of AI chat interfaces?. Feedback hasn't gone away either. A large review of human-GenAI design treats "model feedback to users" as one of six core dimensions How should designers structure human-GenAI interaction systems?. Human Factors researchers borrow "continuous feedback" from automation design as a fix for GenAI productivity losses, though they say plainly that these proposals are still untested Can Human Factors design fixes restore GenAI productivity losses?.

The less obvious point is that AI systems are also starting to run their own kind of feedforward on *us*. Instead of the interface previewing its output for the user, the system reads the user before they submit anything. One system infers your immediate goal by watching your screen, and its outputs are preferred 66–86% of the time over a standard LLM's Can systems infer what users want from watching their screen?. Other work treats gaze, hesitation and typing speed as live signals of what you're thinking Can AI systems read cognitive state from interaction patterns alone?. The same research flags the catch: the signals that let a system time its help well also let it profile and manipulate you. So the real design question goes beyond feedforward versus feedback. Who gets to anticipate whom, and whether the user can see that anticipation happening, matter just as much.

The collection has one paper built squarely on feedforward for AI. The surrounding material comes at it from the side, through chat critiques, context design and intent inference, rather than through the classic HCI literature on feedforward.