Implicit Feedback Signals
What you'll learn: How to gather quality signals from user behavior without asking users to explicitly rate or review outputs.
While explicit feedback (thumbs up/down, ratings) is valuable, most users won't provide it. Implicit feedback signals are behavioral patterns you can observe and measure to infer whether an LLM output was useful—even when users say nothing.
Think of it like a restaurant owner watching diners: if someone devours their meal and licks the plate, that's better feedback than a comment card. Similarly, if a user copies an LLM's response to their clipboard, that action implies the output was valuable.
Common Implicit Signals
Copy actions are gold. When users copy text, they're voting with their behavior that the output is worth using.
Regeneration requests signal dissatisfaction. If someone immediately clicks "regenerate," the first attempt probably missed the mark.
Session duration reveals engagement. A user who stays in conversation for 10 minutes is likely finding value; one who abandons after 15 seconds may not be.
Edit distance (if you track it) shows how much users modify the output. Minor tweaks suggest the LLM got close; complete rewrites suggest it didn't.
Follow-up questions indicate whether the output answered the original query. Immediate clarification requests may signal confusion or incompleteness.
Implementation Approach
After designing your in-app feedback mechanisms (lesson 859), instrument your application to log these behaviors quietly in the background. Aggregate them alongside any explicit feedback you receive, creating a richer picture of model performance in production. These signals become training data for continuous improvement and can trigger alerts for quality degradation.
Key Takeaway: Implicit feedback signals like copy actions, regenerations, and session duration provide scalable quality indicators that capture real user satisfaction without requiring explicit input.