Large Language Models Help Humans Verify Truthfulness -- Except When They Are Convincingly Wrong