Investigating the Perception of the Future in GPT-3, -3.5 and GPT-4
Diana Kozachek
Abstract
This study explores the potential of GPT-3, GPT-3.5, and GPT-4, in generating human-like future scenarios to investigate each model's ability to perceive time. The methodology combines a coding-based experiment and an expert survey. The investigation involves fine- and prompt-tuning GPT-3, prompt-tuning GPT-3.5, and few-shot prompting GPT-4 with human-made future scenarios. The models and output are quantitatively and qualitatively analyzed. The survey invited practitioners from fields of foresight and futurology, AI, and NLP to assess whether differences in output can be identified. This study's findings suggest that GPT-3 and GPT-4 generated scenarios are difficult to distinguish from human-made ones, while GPT-3.5 performed more poorly. Yet none of the models can differentiate time horizons and their respective effects on the future from each other. And while no one knows the shape of things to come, this lack of understanding of a core concept of life invites future investigations.