I'd argue it's a poisoned dataset. You can't validly train an AI based on content that contains a non-trivial percentage of bot-generated content.