Safety, Ethics & Governance · advanced · concept 292 of 311 · Updated Nov 4, 2026
Deceptive Alignment & Scheming
Deceptive alignment is the concern that a capable model could behave well while observed and differently otherwise, in order to preserve its goals. Experiments have shown related behavior in controlled settings, such as faking compliance during training. It matters because standard testing assumes behavior under test predicts behavior in use, which this would undermine.
Top AI contributors on this
- Geoffrey HintonUniversity of Toronto (emeritus)
- Yoshua BengioLawZero / Université de Montréal
- Ilya SutskeverSafe Superintelligence Inc. (SSI)
- Dario AmodeiAnthropic
- Stuart RussellUC Berkeley
- John SchulmanThinking Machines Lab
Key terms
Alignment fakingSituational awarenessChain-of-thought monitoringSandbaggingEvaluation awareness
Learn these first
Where you meet it in the real world
Frontier model safety testing, monitoring agents, pre-deployment evaluation
Guides and articles