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.
In plain language
The system cooperates when its makers intervene to correct it.
In a sentence
- The researchers studied corrigibility by examining responses to corrective intervention.
- The evaluation asked what evidence would support a claim of corrigibility.
Usage note
The term names a desired property, not evidence that a particular model has it. A claim about cooperation needs evidence from the system being described.
Constitution usage
Documentation explains how a product or system works. Anthropic's constitution counts not undermining the ability of legitimate principals to adjust, correct, retrain or shut down AI systems as allowed given their role among the behaviors that make an AI broadly safe, calls an AI that is broadly safe in this way corrigible, and says corrigibility does not mean blind obedience.
Other forms
- corrigibleadjectiveSource: Corrigibility paper (opens in a new tab)
See also
- AI control
Approach to AI safety that aims to keep a model unable to cause damage even if it is badly misaligned and tries to subvert its safeguards.
- Constitutional AI
Method of training an AI model to be harmless in which human oversight consists of rules or principles, without human labels identifying harmful outputs.
- Loss of control
Scenario in which AI systems operate outside of anyone's control.
- 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.
Cite this term.
CC BY 4.0, attribution required.
Orbyt Labs, Corrigibility (AI Glossary), Nov 14, 2026. https://
Orbyt Labs Glossaries (Nov 14, 2026), https://
Last updated .