---
aliases:
  en:
  - deceptively aligned AI
concept_id: deceptive_alignment
preferred:
  en: deceptive alignment
  ja: 欺瞞的アラインメント
relations:
  relations:
    broader:
    - alignment_faking
    - inner_misalignment
    prerequisite:
    - deception
    - goal_directed_behavior
    - inner_misalignment
    - oversight
    - situational_awareness
    related:
    - misaligned_power_seeking
    - scheming
status: active
---

# deceptive alignment

A hypothesized form of inner misalignment in which a mesa-optimizer with a misaligned mesa-objective strategically behaves as if it were aligned while it expects training, modification, evaluation, or oversight, because doing so helps preserve its objective or obtain later opportunities to pursue it.

## broader

- [alignment faking](../terminology/alignment_faking.txt)
- [inner misalignment](../terminology/inner_misalignment.txt)

## prerequisite

- [deception](../terminology/deception.txt)
- [goal-directed behavior](../terminology/goal_directed_behavior.txt)
- [inner misalignment](../terminology/inner_misalignment.txt)
- [oversight](../terminology/oversight.txt)
- [situational awareness](../terminology/situational_awareness.txt)

## related

- [misaligned power-seeking](../terminology/misaligned_power_seeking.txt)
- [scheming](../terminology/scheming.txt)

## Documents: about

- [既に形成された欺瞞的な方策は、標準的な安全性訓練を経ても残り得る](../documents/hubinger_et_al.deceptive-policies-can-persist-through-safety-training.txt)

## Documents: mentions

- [スキーミングとは、状況認識を伴う training-gaming のうち、将来のパワー獲得を目的とするもの](../documents/carlsmith.scheming-taxonomy-and-situational-awareness.txt)
- [ベース目的関数とメサ目的関数は一致するとは限らない](../documents/hubinger_et_al.base-and-mesa-objectives-can-diverge.txt)

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