The paper points out several problems with these large language models. The input coming in is biased because it's trained on huge data sets like all of English Wikipedia, like all of Reddit. So the product that they have created called chat GPT has some guardrails that have been put on it. But right now there is lots of problems with bias as well as misinformation that can come from it.