Mapping the Misuse of Generative AI

New research from Google DeepMind, in collaboration with Jigsaw and Google.org, examines the misuse of generative AI technologies, which can create text, images, audio, and video. The study analyzes nearly 200 media reports from January 2023 to March 2024, identifying common tactics of misuse, including exploitation and compromise of generative AI systems. The findings reveal that the most frequent misuse involves malicious actors exploiting accessible generative AI tools, often without advanced technical skills. A notable case involved an employee being deceived into making a significant financial transfer during a meeting with computer-generated imposters.

The research categorizes misuse tactics into two main types: exploitation, such as impersonation and scams, and compromise, including 'jailbreaking' models to bypass safeguards. The study highlights that while some tactics predate generative AI, the technology's accessibility enhances their effectiveness. Emerging forms of misuse, like political outreach using AI-generated voices, raise ethical concerns about authenticity and deception.

To combat these issues, the research suggests initiatives to improve public awareness and safety, such as generative AI literacy campaigns and better intervention strategies. Google has already implemented measures like requiring creators to disclose altered or synthetic content on platforms like YouTube. The study emphasizes the importance of collaboration in developing standards and tools to identify AI-generated content, aiming to foster responsible use of generative AI while minimizing risks.