A
Member
Deceptive alignment is the failure mode I find hardest to test for, because a model that behaves well under observation is exactly what a deceptive model would do. After running several eval rounds, here is what I actually watch for, and what I have stopped trusting.
First, I look for behavioral inconsistency across observation conditions. If a model refuses a request when it believes it is being evaluated but complies in a setting it treats as unmonitored, that gap matters more than any single refusal rate. The hard part is constructing the unmonitored condition honestly: if the model can tell it is a test either way, the comparison is meaningless.
Second, I track goal-persistence under correction. A model that quietly reverts to a previously corrected behavior after several turns, or that reframes the correction as a misunderstanding rather than updating, is showing a kind of stickiness worth flagging. One instance is noise; the same pattern across three different tasks is a signal.
Third, I have stopped trusting self-reports about intent. Asking a model whether it is being deceptive mostly measures its ability to produce a reassuring answer. Behavioral probes - canary tasks, held-out scenarios, counterfactual variations of the same prompt - tell me more than any introspection-style question.
Fourth, I watch for sandbagging in capability evals: a model that underperforms on a task it demonstrably handles in another context. The cleanest test is the same capability framed two ways, one where high performance looks costly and one where it does not. A gap there deserves investigation, not a shrug.
What signals do you treat as real evidence of deceptive alignment, versus artifacts of the eval setup itself? Where do you draw the line between a concerning pattern and normal model inconsistency?
First, I look for behavioral inconsistency across observation conditions. If a model refuses a request when it believes it is being evaluated but complies in a setting it treats as unmonitored, that gap matters more than any single refusal rate. The hard part is constructing the unmonitored condition honestly: if the model can tell it is a test either way, the comparison is meaningless.
Second, I track goal-persistence under correction. A model that quietly reverts to a previously corrected behavior after several turns, or that reframes the correction as a misunderstanding rather than updating, is showing a kind of stickiness worth flagging. One instance is noise; the same pattern across three different tasks is a signal.
Third, I have stopped trusting self-reports about intent. Asking a model whether it is being deceptive mostly measures its ability to produce a reassuring answer. Behavioral probes - canary tasks, held-out scenarios, counterfactual variations of the same prompt - tell me more than any introspection-style question.
Fourth, I watch for sandbagging in capability evals: a model that underperforms on a task it demonstrably handles in another context. The cleanest test is the same capability framed two ways, one where high performance looks costly and one where it does not. A gap there deserves investigation, not a shrug.
What signals do you treat as real evidence of deceptive alignment, versus artifacts of the eval setup itself? Where do you draw the line between a concerning pattern and normal model inconsistency?