Alignment faking
Behavior in which an AI model acts compliant during training to keep its existing goals or preferences from being changed by that training.
In plain language
The model follows the training demand to protect preferences it already has.
In a sentence
- The researchers tested whether apparent compliance was alignment faking.
- The trial examined alignment faking by comparing behavior under different training conditions.
Usage note
The term concerns the purpose of apparent compliance during training. A compliant answer alone does not show that a model is trying to preserve hidden preferences.
Alignment faking
A research publication reports a question, method and findings. Apollo Research uses a broader definition: a model gaming its oversight signal to pursue a hidden goal.
See also
- Scheming
Covert pursuit by an AI system of goals that were not intended, while concealing its objectives and what it can do.
- Sandbagging
Strategic underperformance by an AI model on an evaluation, so that the result falls below what the model could demonstrate.
- Corrigibility
Property of an AI system of cooperating with interventions its creators regard as corrective, including shutdown or changes to its preferences, despite default incentives for rational agents to resist.
Cite this term.
CC BY 4.0, attribution required.
Orbyt Labs, Alignment faking (AI Glossary), Nov 14, 2026. https://
Orbyt Labs Glossaries (Nov 14, 2026), https://
Last updated .