---
aliases:
  en:
  - hard left turn
concept_id: sharp_left_turn
preferred:
  en: sharp left turn
  ja: シャープ・レフト・ターン
relations:
  relations:
    broader:
    - alignment_failure
    prerequisite:
    - capability_generalization
    - distribution_shift
    - generalization
status: active
---

# sharp left turn

A hypothesized transition in which an AI begins to generalize its capabilities far beyond its training distribution into strategically important domains while its alignment or corrigibility properties fail to generalize comparably, causing earlier safety assumptions to break. In Nate Soares's usage, the key feature is this asymmetry in generalization, not recursive self-improvement or a generic discontinuity by itself.

## broader

- [alignment failure](../terminology/alignment_failure.txt)

## prerequisite

- [capability generalization](../terminology/capability_generalization.txt)
- [distribution shift](../terminology/distribution_shift.txt)
- [generalization](../terminology/generalization.txt)

## Documents: about

- [能力が訓練分布の外へ急速に汎化する局面で、アラインメント特性は同程度には汎化しない可能性がある](../documents/soares.sharp-left-turn.txt)

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