Collecting signal from users about whether the output was any good, and actually using it. Thumbs up and down are nearly free to add and produce weak, heavily biased data. The stronger signals are behavioural: what people edited, what they copied, what they abandoned.
Technical view
Explicit feedback suffers from very low response rates and selection bias towards extremes. Implicit signals such as edit distance between generated and final text, copy events, regeneration, abandonment and downstream task success are more plentiful and more representative. A usable pipeline routes negative signals into a triage queue, produces evaluation cases from them, and closes the loop by reporting whether the fix worked.
Related
Part of AI Terminology Index, section: Product and Adoption.