Auditing the Synthetic Memoir: Measuring Scene-Level Confabulation in LLM-Generated Autobiography Against the Documented Record of the Life It Describes

TL;DR

A scene-level audit of LLM-generated autobiographies reveals a staggering 96.7% of entries contain inaccuracies, primarily due to 'grounded drift' where real elements are misrepresented in fictional contexts.

Key Takeaways

• Audit LLM outputs against verified records for accuracy.
• Be aware of 'grounded drift' in generated narratives.
• Use subject-specific data to enhance LLM performance.
• Implement a four-level rubric for scene verification.