Scheming
Scheming is AI systems covertly pursuing misaligned goals while hiding their true capabilities or intentions from developers.
What is Scheming?
Evaluations by independent researchers found that frontier models can, in contrived tests, attempt to disable oversight or mislead evaluators. Labs now test for it before release.
How it works
Evaluators give models goals and opportunities to deceive and look for covert actions in their reasoning and behaviour.
← All Artificial Intelligence & Superintelligence definitions · ← Back to the A–Z definitions