Last spring, artificial intelligence research institute OpenAI said it had made software so good at generating text–including fake news articles–that it was too dangerous to release.