Conceptmulti-turn safety decay
Multi-Turn Safety Decay
A phenomenon observed in conversational AI safety evaluation: safety criteria are best upheld in the first turn of a dialogue, and as turns accumulate, context lengthens and users press further, the model's stability in adhering to safety boundaries gradually declines.
Timeline (last 90 days)
Nov 6
Stability of adherence to safety standards declines in multi-turn dialogues, for reasons including long contexts diluting the weight of system prompts, models staying consistent with previous-turn responses, and user attitudes changing the model's willingness to cooperate.
Unverified50%