Identifying Life-Changing Books with LLMs (2024) (blog.joellehman.com)

Researchers used LLMs to scan over six million Goodreads reviews (covering ~24,000 books, dataset through 2017) and produce a ranked list of the 300 books most often described by readers as having "changed their life." The pipeline embeds review sentences with an OpenAI embedding model, then applies a trained classifier to flag sentences that express life-changing impact. For each book the system computes the fraction of reviews containing that sentiment and ranks books by that score (the top score is ~0.079, i.e., ~7.9% of reviews flagged). The output is an interactive table (search, sort, genre filter) plus analyses like most life-changing authors and validation tests described in the post. This work demonstrates that LLMs can quantify subjective, transformative experiences at scale, offering a new signal for discovery and recommendation systems that goes beyond star ratings or popularity. Key implications: it surfaces different books (self-help and spiritual titles dominate the top entries) than popularity/top-rated lists, enabling personalization around impact rather than hype. Important caveats include dataset age and sampling bias, reliance on self-reported language (which can vary by culture and genre), and classifier/embedding errors — all of which affect fairness and representativeness. Still, the approach is a practical example of using embeddings + supervised classification to extract nuanced human judgments from large text corpora.