Large language models are able to downplay their cognitive abilities to fit the persona they simulate