Glossary
Misalignment
A mismatch between an AI system’s behaviour or objectives and the human intentions or requirements it should serve.
In the book
Diagnose the actual problem. Missing information, an ignored instruction and an invented fact call for different responses. A disappointing answer does not automatically show misalignment, and a model can make mistakes even after a clear brief.
Introduced in chapter 3, “ResponsAIbility”.
Related terms
- AI alignment: The effort to make AI systems behave in accordance with intended human goals, values and constraints.