Verbosity decreases accuracy in large language models