Large language models are able to downplay their cognitive abilities to fit the persona they simulate
Fig 5
Proportion of correct answers by child’s gender and age.
Charts on the left represent results generated by GPT-3.5-turbo, and charts on the right represent results generated by GPT-4. The first line of charts represents the plain zero-shot prompt, the second chain-of-thoughts prompt, and the third represents the primed-by-corpus prompt.