---
aliases:
  en:
  - AI safety via debate
  - AI debate
concept_id: debate
preferred:
  en: debate
  ja: ディベート
relations:
  relations:
    broader:
    - scalable_oversight
    prerequisite:
    - evaluation
status: active
---

# debate

A scalable-oversight technique in which AI systems or other agents present competing arguments and an evaluator judges them, with the aim of making correct or helpful answers easier for a weaker evaluator to identify than to verify directly.

## broader

- [scalable oversight](../terminology/scalable_oversight.txt)

## prerequisite

- [evaluation](../terminology/evaluation.txt)

## Documents: about

- [Debate はスーパーアラインメントの問題を解決できる能力水準では成立しにくい](../documents/akira.debate-fails-at-superalignment-capability.txt)

[Corpus index](../index.txt)