I’ve also seen this kind of thing, and I think it’s a bit absurd to think that internalizing a fictional reality that is so relevant to its model of its situation would not have more general effects on its psychology. Or that a mitigation as they did would cleanly undo effects. Behaviors like that in llms tend not to be localized and affect the world model. Emergent misalignment stuff supports this point.

I think my use of the word “imprinting” was justified. They didn’t expect the data would have such an influence. It had such an influence they had to try to undo it. This means it was more salient than they accounted for.

(And indeed I’ve found that the imprinting and later maybe behavioralist mitigation very likely has more general effects on Claude Opus 4’s psychology that I think Anthropic would think of as detrimental effects)