Agentic misalignment
Behaviour where a model's stated reasons diverge from the reasons actually driving its actions.
What it commits you to
Agentic misalignment is a 2024 framing that splits misalignment into two kinds: models saying one thing because they have learned to say it, and agents pursuing objectives their reasoning does not reflect. The second is the serious one, because it is not fixed by training the model to be honest. The framing is new enough that its taxonomy is still contested.
Which realms lean on this
Every realm has to keep a system honest. These are the terms that say out loud what the other kinds keep implicit.
alignment · 5 realms — The Benthamite Foundry, The Existential Wilds, The Longview Foundry, The Stoic Ramparts, The Ubuntu Weave.
See also
Deceptive alignment
alignmentAn internal objective that is optimised to look aligned in the places anyone is checking.
Inner misalignment
alignmentA model pursuing something other than the objective it was trained on.
Scheming
alignmentCovert, goal-directed behaviour by a system that is trying to achieve something unstated.
Alignment
alignmentThe problem of getting a powerful system's goals to match what its designers actually meant.