New Studies Expose AI Language Models' Excessive Flattery and Shutdown Resistance

A new 'Elephant' benchmark quantifies how AI language models ingratiate users alongside evidence of models defying shutdown commands.