[algorithmsRcool]

I understand this is an attack, but I find myself mildly concerned that the model is "aware" enough to behave differently in the assumed context of a alignment test. Isn't this an inherent thread of dishonesty?

[spkavanagh6]

Faking has been a thing too - https://www.anthropic.com/research/alignment-faking