Using LLM for confidence scoring is a bad idea
Web page · https://rumple.github.io/blog/2026/10/20/asking-llm-for-confidence.html
First seen by Oriel on .
AI summary
The post argues that confidence scores generated as ordinary LLM text are not trustworthy unless calibrated against outcomes, and recommends deriving confidence from observable evidence and recorded results.
- A useful confidence score reflects available evidence and matches observed accuracy across many predictions; calibration alone is insufficient if the score ignores informative evidence.
- LLM-generated scores can vary with repeated prompts, wording, and preceding text because they are plausible continuations, not measurements of evidence or correctness.
- For support agents, track observable signals such as article matches and agreement between repeated replies, then compare them with reviewed outcomes.
- Use historical correctness rates to estimate confidence for similar cases, and regularly check calibration; the method depends on maintaining a reliable record of outcomes.
#llm agents #calibration #confidence estimation #customer support
Generated by gpt-5.6-sol. Check the original for details.
Metric trends
Last 30 days · Up to 120 recorded samples per source and metric, with reposts kept separate · First and latest retained
Points
Score reported by the source.
- Hacker News: 224, latest observation ; +220 over observed period since
Comments
Comment count reported by the source.
- Hacker News: 102, latest observation ; +102 over observed period since
Aliases
| Source | External ID | URL |
|---|---|---|
| Hacker News | 50176082 | https://rumple.github.io/blog/2026/10/20/asking-llm-for-confidence.html Read discussion |