Entering the age of hybrid futures: A comparative study of human and GPT-generated scenarios

The increasing use of Large Language Models (LLMs) in scenario generation presents new opportunities and challenges for futures research. While scenario development has traditionally emphasized human creativity, judgment, and deliberation, generative AI models such as GPT-3, GPT-3.5, and GPT-4 are now capable of producing seemingly plausible futures narratives at scale. This study investigates how scenarios generated by three generations of GPT models compare to human-authored scenarios on Europe’s futures. Using a curated dataset of 1,080 human-authored scenarios pre-ChatGPT, the models were fine-tuned and prompt-engineered to generate 3,240 scenarios. A comparative analysis was conducted using quantitative measures (n-gram overlap, lexical diversity, topic modeling, sentiment analysis) and qualitative expert evaluation through a Delphi survey (N = 42). Results reveal substantive differences in thematic emphasis, linguistic features, and emotional tone between human and machine-generated outputs. Experts frequently misclassified the origin of scenarios, highlighting the increasingly blurred boundary between human and AI authorship. However, plausibility assessments showed a subtle bias against scenarios labeled as AI-generated, while AI-generated scenarios were thematically biased towards technology-centered topics. These findings raise important questions for the methodological integration of LLMs into scenario practices and suggest future directions for hybrid human–AI collaboration in anticipation research. This study contributes to the notion that scenario work increasingly relies on hybrid human–AI collaboration, requiring futures researchers to engage with technology-augmented methods in meaningful ways. • Investigates potentials and challenges of GPT-3, GPT-3.5, and GPT-4 for Futures Research scenarios. • Methodology combines coding-based experiment, technical evaluations, and expert survey. • Analyzes outputs quantitatively and qualitatively, bridging objective and subjective evaluations. • Surveyed 42 practitioners to assess content differences and scenario plausibility. • GPT models heavily center on technology-driven topics, struggle with time-horizon distinction. • Experts couldn't differentiate GPT-generated scenarios from human-made ones, except GPT-3.5. • Comparing human-made to GPT-generated scenarios remains complex, implying hybrid collaboration. • Discusses implications and weaknesses of GPT-3, GPT-3.5, and GPT-4 for Futures Research. • Provides practical insights for further research and publicly accessible process, including code and corpus.