Without knowledge of how to optimize answers for ‘truth,’ they’re modeling what humans do–tell stories, hedge, prevaricate, lie, do bad math, and sometimes, eventually, suss out the truth.