AI CoT may work the same extremely flawed way that human introspection does, and that’s fine, the reason we may want to hold them to a higher standard is because someone proposed to use CoTs to monitor ethics and alignment.