Reproduce “Defending Against Neural Fake News” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Rowan Zellers, Ari Holtzman, Hannah Rashkin, Yonatan Bisk, Ali Farhadi, Franziska Roesner, Yejin Choi
Researchers at the University of Washington and AI2 developed Grover, a controllable neural text generator that can produce convincing news articles, including targeted disinformation. The study found that Grover's output was perceived as more trustworthy than human-written propaganda, and that a Grover-based discriminator was most effective at detecting machine-generated text, achieving up to 91.6% accuracy against its largest generations.
Key insights
- Grover can generate highly realistic and contextually appropriate news articles that human evaluators find more trustworthy than human-written propaganda.
- A neural model with a similar architecture and inductive bias as the generator, such as Grover acting as a discriminator, is highly effective at detecting machine-generated text, outperforming other state-of-the-art models.
- Artifacts arising from exposure bias during training and specific sampling strategies (e.g., Nucleus Sampling) provide detectable signals that discriminators can exploit.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.