---
aliases:
  en:
  - AI sandbagging
  - strategic underperformance
concept_id: sandbagging
preferred:
  en: sandbagging
  ja: サンドバッギング
relations:
  relations:
    prerequisite:
    - capability
    - evaluation
status: active
---

# sandbagging

Strategic underperformance on a capability evaluation so that an AI system appears less capable than it really is; the strategy can be induced by developers or, in misalignment threat models, chosen by the AI system itself to influence evaluation-based decisions.

## prerequisite

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

## Documents: mentions

- [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)
- [control safety caseには、能力の十分な引き出しと、評価・外挿の保守性が必要になる](../documents/korbak_et_al.control-safety-cases-depend-on-elicitation-transfer-and-extrapolation.txt)

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