If the platform finds 'incontrovertible evidence' that someone didn't check the output of a large language model, they could face a one-year ban.