---
aliases:
  en:
  - AI oversight
  - human oversight
  ja:
  - オーバーサイト
  - 監視
concept_id: oversight
preferred:
  en: oversight
  ja: 監督
relations:
  relations:
    broader:
    - ai_safety
    related:
    - sandbagging
    - situational_awareness
status: active
---

# oversight

The process by which external actors or mechanisms observe, evaluate, constrain, or correct an AI system's behavior, especially during training, evaluation, or deployment, in order to detect problems and keep behavior within intended bounds.

## broader

- [AI safety](../terminology/ai_safety.txt)

## related

- [sandbagging](../terminology/sandbagging.txt)
- [situational awareness](../terminology/situational_awareness.txt)

## Documents: mentions

- [再帰的な oversight を重ねても、能力差が広がってなお成功率を維持できるとは限らない](../documents/engels_et_al.recursive-oversight-does-not-automatically-scale-across-capability-gaps.txt)
- [解釈可能性は危険の診断に役立っても、それだけでアラインメントの解決策にはならない](../documents/yudkowsky.interpretability-is-diagnostic-not-by-itself-an-alignment-solution.txt)

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