Rethinking Storytelling: Human vs AI Narratives - A Cultural Perspective

Mike Young - Nov 6 - - Dev Community

This is a Plain English Papers summary of a research paper called Rethinking Storytelling: Human vs AI Narratives - A Cultural Perspective. If you like these kinds of analysis, you should join AImodels.fyi or follow me on Twitter.

Overview

  • This paper proposes a framework that combines behavioral and computational experiments using fictional prompts to investigate cultural artifacts and social biases in storytelling by both humans and generative AI.
  • The study analyzes 250 stories authored by crowdworkers in 2019 and 80 stories generated by GPT-3.5 and GPT-4 in 2023, using methods from narratology and inferential statistics.
  • The experimental paradigm allows for a direct and controlled comparison between human and large language model (LLM) generated storytelling.

Plain English Explanation

The researchers wanted to understand the cultural beliefs and biases reflected in stories written by both humans and AI language models. To do this, they created a set of prompts that asked people and AI models to write stories about falling in love with an artificial human. By looking at the stories that were produced, the researchers could see what kinds of ideas and assumptions were present in the collective imagination of both humans and AI.

The researchers gathered 250 stories written by crowdworkers in 2019 and 80 stories generated by the GPT-3.5 and GPT-4 language models in 2023. They used a combination of narrative analysis and statistical methods to compare the stories and identify patterns.

The key idea is that fiction can be a window into the beliefs and social dynamics that shape how both humans and AI systems think about the world. By using a controlled experimental setup with the same prompts, the researchers were able to draw direct comparisons between the human and AI-generated stories.

Key Findings

  • The narratives, whether written by humans or generated by AI, all depicted a scientific or technological pursuit, reflecting the pervasiveness of the Pygmalion myth in the collective imagination.
  • The stories generated by GPT-3.5 and particularly GPT-4 were more progressive in their portrayal of gender roles and sexuality compared to the human-authored stories.
  • While AI-generated narratives can sometimes offer innovative plot twists, they generally provide less imaginative scenarios and rhetoric than the human-authored texts.

Technical Explanation

The researchers designed an experimental framework that combined behavioral and computational techniques to investigate cultural artifacts and social biases in storytelling. They gathered 250 stories written by crowdworkers in June 2019 and 80 stories generated by the GPT-3.5 and GPT-4 language models in March 2023.

All participants were given the same Pygmalionesque prompts about creating and falling in love with an artificial human. The researchers then used methods from narratology and inferential statistics to analyze the resulting stories.

The key insight is that this experimental paradigm allows for a direct and controlled comparison between human and LLM-generated storytelling. By using the same prompts, the researchers were able to identify patterns and differences in how humans and AI systems conceptualize and depict certain social and cultural elements through narrative.

Implications for the Field

This research demonstrates how fiction can be used as a window into the collective imaginary and social dimensions of both humans and AI systems. The findings suggest that language models like GPT-4 may be more progressive in their representation of gender and sexuality compared to human-authored narratives, though they still lack the imaginative depth of human storytelling.

This work highlights the value of using controlled experimental setups and interdisciplinary methods to investigate the interplay between technology, culture, and social biases. The proposed framework offers a novel approach for studying these complex relationships through the lens of storytelling.

Critical Analysis

The paper acknowledges some limitations, such as the relatively small sample sizes and the focus on a single, specific prompt. Additionally, the researchers note that the AI-generated stories were based on default settings, without any additional prompting or fine-tuning, which may have impacted their level of creativity and imagination.

Further research could explore the impact of different prompting techniques, larger datasets, and more diverse narrative genres to gain a more comprehensive understanding of the relationships between human and AI-generated storytelling. It would also be valuable to investigate how these findings may vary across different cultural contexts and time periods.

Conclusion

This paper presents a novel experimental framework that combines behavioral and computational methods to study the cultural artifacts and social biases reflected in storytelling, both by humans and generative AI systems. The key finding is that while AI-generated narratives can be more progressive in their representation of gender and sexuality, they often lack the imaginative depth of human-authored stories.

The proposed framework offers a valuable tool for researchers to explore the complex interplay between technology, culture, and social dynamics through the lens of fictional narratives. This work highlights the potential of using interdisciplinary approaches to gain deeper insights into the collective imaginary and social dimensions of both human and AI-based storytelling.

If you enjoyed this summary, consider joining AImodels.fyi or following me on Twitter for more AI and machine learning content.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Terabox Video Player