## Definition **Conversational feedback** is the set of signals — both explicit and implicit — that users of a conversational AI application generate through their messages and actions, and that can be harvested to evaluate model quality, guide development, and personalise the product. Unlike traditional software feedback (star ratings, support tickets), conversational feedback is embedded in the natural flow of dialogue, making it richer but harder to extract. ## Why It Matters for AI Products User feedback is proprietary data, and data is a competitive advantage. A product that ships early, attracts users, and captures their feedback can continuously improve its models in a way that is hard for competitors to replicate — the *data flywheel* effect. Conversational feedback has an especially high value because foundation models can be fine-tuned on it: even implicit signals like error corrections can be converted into preference data for [[RLHF]]. ## Explicit vs Implicit Feedback **Explicit feedback** is solicited directly: thumbs up/down, star ratings, "Did this solve your problem?" It is easy to interpret but sparse — many users do not bother, and those who do may exhibit *leniency bias* (rating positively to avoid friction). **Implicit feedback** is inferred from user behaviour. Because it requires no extra effort, it is abundant; but it is also noisier. ## Taxonomy of Implicit Signals **Natural language signals** (extracted from message content): | Signal | Likely interpretation | |---|---| | Early termination (stops generation mid-stream) | Response is not useful | | "No, I meant…" / rephrasing | Model misunderstood | | Explicit error correction ("Bill is the suspect, not the victim") | Factual error; high-quality preference data | | "Are you sure?" / "Check again" | Distrust or missing detail | | User edits the model's output directly | Strong negative signal on original; edited version = winning response | | Complaints ("That's wrong", "Too long") | Failure signal; cluster by type to prioritise fixes | | Positive sentiment or "Perfect" | Success signal | **Behavioural signals** (derived from user actions): - **Regeneration**: user unsatisfied with the response or wants options; stronger signal on usage-based billing (costs money). - **Conversation length**: positive for companionship apps; may indicate inefficiency in customer support. - **Conversation organisation**: deleting a conversation is a strong negative signal; renaming suggests the content was good but the auto-title was bad. - **Comparative selection**: when two options are shown side by side, the chosen one is a preference signal (used for fine-tuning). ## Feedback Design Principles 1. **Non-intrusive collection**: feedback UI must not disrupt user workflow; it should be easy to ignore. 2. **Collect throughout the journey**: at onboarding (calibration), on failure (downvote, regenerate), when the model is uncertain (show options side by side). 3. **Context alongside the signal**: a thumbs-down alone is not actionable; the preceding 5–10 turns make it debuggable. 4. **Understand biases before acting**: - *Leniency bias*: users avoid negative ratings to escape extra steps. - *Randomness*: users click without reading long side-by-side responses. - *Position bias*: first option gets more clicks regardless of quality. - *Preference bias*: longer responses are preferred even when less accurate. 5. **Avoid degenerate feedback loops**: acting naively on feedback can amplify initial biases (popularity bias, filter bubbles) and teach models to be sycophantic rather than accurate. ## Feedback as Training Data Error-correction messages and comparative selections convert directly into preference pairs `(query, winning_response, losing_response)` for [[RLHF]] or DPO. User-edited outputs provide the same structure: original generation = losing, edited version = winning. This pipeline is what makes user feedback a structural competitive moat for AI products. ## Related - [[RLHF]] - [[AI Application Architecture]] - [[Hallucination]] - [[Prompt Injection]] ## Sources - [[AI Engineering - Chip Huyen]]