This paper presents GROVER, a model for controllable text generation that can generate realistic-looking neural fake news. The authors highlight the dual-use concerns of recent advances in natural language generation, noting that while applications like summarization and translation are beneficial, the same technology could be exploited by adversaries to create targeted propaganda that mimics real news. To address this, the authors develop GROVER, which can generate entire news articles, including headlines, sources, dates, and authors. Humans find these generated articles more trustworthy than human-written disinformation. The authors also investigate the effectiveness of discriminators in identifying neural fake news. They find that the best current discriminators achieve 73% accuracy in classifying neural fake news from real news, but when used as a discriminator, GROVER itself achieves 92% accuracy, demonstrating the importance of public release of strong generators. The authors further explore how exposure bias and sampling strategies affect the generation of neural fake news, and how these factors can be leveraged by discriminators. The paper also discusses the ethical implications of releasing such models and proposes a provisional policy for their release. The authors argue that releasing GROVER is essential for developing better detection methods for neural fake news. They also discuss the challenges of detecting neural fake news, noting that while models like BERT are strong discriminators for many NLP tasks, they are not as effective at detecting GROVER-generated fake news as left-to-right models like GROVER. The authors conclude that GROVER is an effective discriminator when given a medium number of fake news examples from the exact adversary that will be encountered at test time. However, if the assumption is relaxed, the performance of GROVER decreases significantly. The paper also discusses the importance of variance reduction in generating text and how it affects the ability of discriminators to detect neural fake news. The authors suggest that a sweet spot exists for how much variance to reduce, which makes discrimination harder for models. They also note that the most adversarial top-p threshold for BERT is lower than that for GROVER, suggesting that BERT's view of language differs from GROVER's. Finally, the authors discuss the importance of releasing generators for the development of better detection methods and the ethical considerations involved in doing so. They argue that releasing GROVER is essential for creating better defenses against neural fake news. The paper concludes with a discussion of future research directions, including the development of more powerful generators and the integration of knowledge into discriminators for better fact verification.This paper presents GROVER, a model for controllable text generation that can generate realistic-looking neural fake news. The authors highlight the dual-use concerns of recent advances in natural language generation, noting that while applications like summarization and translation are beneficial, the same technology could be exploited by adversaries to create targeted propaganda that mimics real news. To address this, the authors develop GROVER, which can generate entire news articles, including headlines, sources, dates, and authors. Humans find these generated articles more trustworthy than human-written disinformation. The authors also investigate the effectiveness of discriminators in identifying neural fake news. They find that the best current discriminators achieve 73% accuracy in classifying neural fake news from real news, but when used as a discriminator, GROVER itself achieves 92% accuracy, demonstrating the importance of public release of strong generators. The authors further explore how exposure bias and sampling strategies affect the generation of neural fake news, and how these factors can be leveraged by discriminators. The paper also discusses the ethical implications of releasing such models and proposes a provisional policy for their release. The authors argue that releasing GROVER is essential for developing better detection methods for neural fake news. They also discuss the challenges of detecting neural fake news, noting that while models like BERT are strong discriminators for many NLP tasks, they are not as effective at detecting GROVER-generated fake news as left-to-right models like GROVER. The authors conclude that GROVER is an effective discriminator when given a medium number of fake news examples from the exact adversary that will be encountered at test time. However, if the assumption is relaxed, the performance of GROVER decreases significantly. The paper also discusses the importance of variance reduction in generating text and how it affects the ability of discriminators to detect neural fake news. The authors suggest that a sweet spot exists for how much variance to reduce, which makes discrimination harder for models. They also note that the most adversarial top-p threshold for BERT is lower than that for GROVER, suggesting that BERT's view of language differs from GROVER's. Finally, the authors discuss the importance of releasing generators for the development of better detection methods and the ethical considerations involved in doing so. They argue that releasing GROVER is essential for creating better defenses against neural fake news. The paper concludes with a discussion of future research directions, including the development of more powerful generators and the integration of knowledge into discriminators for better fact verification.