A fiction archive for comparing frontier AI model behavior over time (frontierfictionarchive.org)

The Frontier Fiction Archive has been launched to analyze and preserve the output of frontier AI models through longform speculative fiction, offering a new perspective on AI behavior beyond traditional benchmarks. This initiative seeks to explore how models respond to open-ended creative tasks, examining their narrative choices, thematic concerns, and areas of failure over time. Unlike conventional comparisons that often emphasize quick, surface-level competence, the Archive investigates deeper elements such as character development, pacing, and implied values, thereby enriching the understanding of model capabilities. Each story submitted to the Archive will be meticulously documented with details like model identifiers, run dates, and editorial notes to ensure a thorough context for future analysis. The project's intent is to create a longitudinal record that allows for meaningful comparisons between different AI models, assessing their growth in creativity, complexity, and cultural awareness. The team encourages input from a diverse group of stakeholders to refine the project's focus on preserving the intricacies of AI-generated fiction and understanding its evolution, paving the way for better insights into the nature of AI creativity moving forward.