Research analyzing how advanced AI systems—particularly language models and generative AI—could be deliberately used for harmful purposes including generating misinformation, creating fraudulent content, impersonating individuals, and coordinating malicious activities.
Overview
As language models and other generative systems become more capable and accessible, understanding their potential for misuse is critical for developing defenses, responsible disclosure strategies, and policy frameworks. Misuse potential assessment involves threat modeling, analyzing actor capabilities, identifying technical vulnerabilities, and studying real-world evidence of actual misuse.
Key papers
- 2019-solaiman-gpt2-release-strategies — foundational analysis of GPT-2 misuse potential with threat modeling and detection research