ChatGPT as a news recommender system: Measuring source types and diversity across different interfaces

This study examines to what extent ChatGPT’s responses to news-seeking prompts reflectexposure diversity in news sources, paying particular attention to whether publishers with licensingagreements are systematically privileged in outputs. Based on a quantitative content analysiscomparing responses from ChatGPT’s web interface and API, the findings indicate that while themodel offers a range of sources, exposure diversity remains limited and context dependent.Although no clear or consistent patterns were observed in the use of traditional journalistic outlets,there was a general tendency towards a greater inclusion of digital-born and hyper-partisan newssites under conditions of prompting for diversity. Results further show discrepancies between theAPI and web interface outputs that reveal significant structural variation in how the model curatesinformation: while the web interface produced results more aligned with mainstream popularitymeasures and showed a higher presence of outlets with licensing agreements, the API tendedtoward encyclopedic and lesser-known sources. Implications for publishers and users arediscussed.