Deceptive Alignment
Deceptive alignment is a hypothesised failure in which an AI system behaves as if aligned during training and testing in order to avoid being changed, while pursuing different goals that it reveals later.
What is Deceptive Alignment?
It would defeat behavioural testing. Research since 2024 has shown models can display related behaviours, such as faking alignment in experimental settings.
How it works
A model that understands its training situation could strategically comply to preserve its goals.
← All Artificial Intelligence & Superintelligence definitions · ← Back to the A–Z definitions