Feedback Collection That Is Worth Having

Every AI product ships a thumbs up and a thumbs down. The response rate is low single digits, the responders are unrepresentative, and the signal is a binary judgement of an entire output with no indication of which part was wrong or what "wrong" meant. It is the cheapest thing to build and close to the least informative thing to collect.

Implicit signals carry more

Behaviour reveals evaluation without asking for it, and the response rate is one hundred percent.

Acceptance. Whether the user kept the output, copied it, sent it, committed it. This is the strongest single signal available in most products and it is often not instrumented.

Editing. What the user changed before using it. The edit is a correction with a diff, which is far richer than a rating and is directly usable as preference data.

Retry and rephrase. A user who immediately asks again, or asks the same thing differently, has judged the first attempt inadequate. Retry rate is a quality metric with no annotation cost.

Abandonment. Leaving mid-generation or ending the session after an output is a negative signal, though a noisy one since people also leave for unrelated reasons.

Time to next action. A long pause before proceeding suggests the output required work to verify or interpret.

The interpretation caveat is that implicit signals measure satisfaction and acceptance, not correctness. A user who accepts a wrong answer they did not verify has produced a positive signal for a bad output, which is exactly the case that matters most in high-stakes settings.

Making explicit feedback informative

Where explicit feedback is collected, three changes make it worth more than a rating.

Ask at the point of the problem. Feedback attached to a specific sentence, claim or step is actionable; feedback on a whole response is not.

Ask what, not how much. A short list of failure categories, factually wrong, missed the question, wrong tone, too long, unsafe, produces a distribution that tells a team where to work. A star rating produces a number that moves for reasons nobody can decompose.

Capture the correction. The single most valuable artefact is the user's corrected version, because it is a training example rather than a signal. Making editing easy and capturing the edit is worth more than any rating scheme.

When it breaks

Response bias dominates explicit feedback. People rate when they are annoyed or delighted, so the collected distribution is bimodal and unrepresentative of the middle where most outputs sit. Aggregate ratings should not be read as an average quality estimate.

Feedback as training data closes a loop. Training on user-accepted outputs teaches the model to produce what users accept, which correlates with quality and also with confidence, agreeableness and length. Optimising accepted outputs directly is a route to sycophancy, and it needs a counterweight from independent evaluation.

Nobody sees the feedback. Collected feedback routed to a dashboard nobody reads is instrumentation without a consumer. A weekly review of a sample of negative feedback, by someone who can change something, is what converts collection into improvement.

Asking has a cost. Every prompt for feedback is an interruption, and a product that asks after every interaction trains users to dismiss. Sampling, so that a small fraction of interactions get an ask, gets most of the signal for a fraction of the friction.