The thumbs down button is not evil.
It is just wildly incomplete.
Most users do not give feedback. They do not click thumbs down. They do not write a helpful note explaining that the agent misunderstood the second clause, ignored their constraint, and made them feel like they were debugging a stranger’s brain.
They just leave.
Thumbs down feedback is explicit negative feedback on an AI answer. It is useful, but it only captures the tiny slice of users who are frustrated enough, motivated enough, and patient enough to report the problem.
That is not your failure dataset. That is your complaint dataset.
^ the feedback dashboard with 3 thumbs down while 300 users silently bailed
Why users do not click thumbs down
Because they are trying to get a job done.
The user did not open your agent to become a QA analyst. They wanted a refund, a workout plan, a code fix, a lesson, a candidate screen, or a support answer.
When the agent fails, giving feedback is extra work.
Common reasons users skip it:
- They are in a hurry.
- They do not believe anyone reads it.
- They are unsure whether the answer is wrong.
- They feel awkward being negative.
- They already found a workaround.
- They are annoyed and done.
So the absence of thumbs down does not mean satisfaction. It often means nothing.
That is a very rude property for a metric to have.
What should you measure instead?
Measure behavior.
| Conversation signal | What it can reveal |
|---|---|
| Repeated rephrasing | The agent missed intent |
| Abrupt exit | User gave up |
| “Nevermind” | Trust broke |
| Shorter future asks | User downgraded delegation |
| Same question later | Answer did not resolve need |
| Manual escalation | Automation failed silently |
| Polite thanks after friction | Social exit, not satisfaction |
These signals are not perfect either. But they are closer to the real product experience because they do not require the user to volunteer feedback.
You want to know what users do when the agent disappoints them.
Not just what they report.
^ user sending a sixth rephrase instead of clicking the tiny feedback icon
Where thumbs down still helps
Keep it.
Explicit feedback is useful when you pair it with the surrounding conversation.
A thumbs down attached to one answer is okay. A thumbs down attached to the full session is better. You want the path that produced the complaint:
- What did the user ask first?
- Where did the agent misread intent?
- Did the user correct it?
- Did the agent recover?
- Did the user continue?
The rating is the flare. The conversation is the fire.
The feedback stack that works better
Think of agent feedback in layers.
| Layer | Example | What it catches |
|---|---|---|
| Explicit feedback | Thumbs down, rating, comment | Users willing to report |
| Conversation behavior | Rephrase, correction, exit | Silent friction |
| Long-term behavior | Return, downgrade, churn | Trust impact |
| Team review | Human-labeled failure pattern | Root cause |
The mistake is choosing one layer and pretending it tells the whole truth.
Explicit feedback gives you sharp examples. Behavior gives you coverage. Long-term usage tells you whether the experience changed the relationship. Human review gives you meaning.
Together, they become a feedback system. Alone, thumbs down is a tiny flashlight in a large dark room.
TLDR
Thumbs down feedback is not enough because most failed users never click it.
Use it as one signal, not the quality system.
Measure repeated rephrases, abrupt exits, “nevermind” moments, future delegation shrink, and unresolved repeated questions. That is where silent friction shows up.
Agnost helps teams turn these implicit conversation signals into a more honest view of agent quality, because waiting for users to complain is basically product analytics by astrology.
FAQ
Should we remove thumbs up and thumbs down?
No. Keep them, but do not confuse them with representative feedback.
Is implicit feedback reliable?
It is probabilistic. A single signal can be noisy, but repeated patterns across many conversations are very useful.
What is the biggest mistake teams make with feedback?
They optimize for fewer complaints instead of better outcomes. Silence is not success.