Autonomous, Safe, and Auditable AI Agents for Data Journalism Workflows will automate portions of the data journalism workflow while focusing on security and performance concerns that come with scraping tasks. The project will train AI agents to generate pseudocode instead of code, returning higher-level instructions that are more interpretable, understandable, and secure for users. The AI agent would only work with provided instructions, eliminating the ability to generate arbitrary code. The proposal aims to develop this approach through an MVP, investigating its potential benefits in data journalism workflows and exploring user reactions, deployment models, as well as ethical, legal, and security issues.
Léopold Mebazaa, `16 BA, Columbia University