---
aliases:
  en:
  - outer alignment failure
  - reward misspecification
concept_id: outer_misalignment
preferred:
  en: outer misalignment
  ja: アウター・ミスアラインメント
relations:
  relations:
    broader:
    - misalignment
    prerequisite:
    - misalignment
    - outer_alignment
    related:
    - proxy_objective
status: active
---

# outer misalignment

A failure in which the training objective, reward process, loss, or other specification does not adequately represent the designers' intended target, so optimizing it can produce undesirable outcomes even if the learned system faithfully pursues that specification.

## broader

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

## prerequisite

- [misalignment](../terminology/misalignment.txt)
- [outer alignment](../terminology/outer_alignment.txt)

## related

- [proxy objective](../terminology/proxy_objective.txt)

## Documents: about

- [測定しやすい代理指標の最適化により、人間は単一の支配シフトを経ずとも徐々に制御を失い得る](../documents/christiano.optimizing-measurable-proxies-can-cause-gradual-disempowerment.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)