FG—01AI Design Field Guide

Search the field guide

Search chapters, principles, patterns, worksheets, and glossary terms

PTN—40

Feedback loops

Collect user feedback at the moment of reaction, then visibly change behavior because of it.

Every AI product ships thumbs up/down. Almost none close the loop. The buttons collect training data for a model retrain months away, the user sees nothing change, and within a week they stop clicking. A feedback loop is only a loop if the person can perceive the return path: the correction takes effect now, in this session, on this kind of output — or the interface says honestly that it will take time and where it went.

The strongest feedback mechanisms don’t look like feedback mechanisms. An edit to generated text, a rejected suggestion, a regenerate with a steering note — these are corrections captured in the flow of work, with far richer signal than a binary rating. Cursor and Copilot learn more from which completions get accepted than any thumbs widget could gather, because acceptance is feedback people give without deciding to give it.

When you do ask explicitly, ask small and specific. “Was this helpful?” after every response is a tax; a targeted “Too long?” chip after a visibly long answer is a conversation. And whatever you collect, spend some of it locally: if someone downvotes a verbose answer and the very next answer is shorter, they’ve learned that feedback matters here. That lesson is worth more than the data point.

When to use

Design considerations

Pitfalls

In the wild

GitHub Copilot
Treats completion acceptance and rejection as its core quality signal — feedback captured entirely inside the act of coding, no widget required.
Cursor
Accept/reject on inline diffs makes every AI edit a feedback event; rejected changes vanish cleanly, so saying “no” costs nothing.
ChatGPT
Pairs thumbs-down with a quick reason picker, and regenerate offers a comparison between responses — turning a rating into a preference judgment.
Midjourney
The upscale/variation grid is a feedback loop as the primary interface: choosing one image over three siblings is a preference signal that immediately shapes the next generation.

Principles served

Discussed in the guide