Generative Models¶
Generative models learn the underlying distribution of data and can generate new samples that resemble training data. In the context of misinformation research, the focus is primarily on language models (transformers trained with causal masking), which generate text token-by-token.
The capabilities of generative models for creating synthetic content—both beneficial (code generation, content creation) and harmful (misinformation, deepfakes)—make them central to contemporary discussions of information ecosystem risks.
Key papers¶
- FLIRT: Feedback Loop In-context Red Teaming — Automated red teaming of text-to-image and text-to-text generative models; demonstrates vulnerabilities in Stable Diffusion and language models through adversarial prompt generation
- AIGC Survey — comprehensive overview of generative AI history, techniques, and applications from GANs through transformers to ChatGPT
- DeepFakes and Beyond: A Survey of Face Manipulation and Fake Detection — survey of GAN-based face synthesis and manipulation techniques (StyleGAN, ProGAN)
- Disinformation Capabilities of Large Language Models — evaluation of disinformation generation via instruction-tuned language models
- Zellers et al. (2021) — GROVER: A State-of-the-Art Open-Source Neural Fake News Generator
Related topics¶
- Large Language Models — the primary form of generative model studied in misinformation contexts
- Disinformation Generation — harmful applications of generative models