I do not use large language model AIs (LLMs) to generate indexes, nor do I upload or share client documents with LLMs without express permission.
LLMs fail at the indexer’s primary task: to ensure readers find needed information. Tests I have done show that LLMs underindex book-length works, do not provide adequate structure or cross-references, and can insert false information (hallucinations) into the index.* Each of these is problematic:
- Severe under-indexing, with as few as 20-40% of the access points of human-generated indexes, can prevent readers from locating desired information and mislead them into thinking that omitted information isn’t in the book at all
- Absent/near-absent index structure, especially cross-references, prevents the reader from effectively navigating to subtopics and related topics while misrepresenting the focus of the book
- Hallucinations (including invented page references and even wholly nonexistent topics) waste the reader’s time and break the trust between the reader and the book
Future developments in AI may bring improvements, but at present the human brain of a professional indexer is still the best tool for analyzing, writing, and editing an index in full awareness of context as per quality standards.
I do use LLMs when developing software utilities, as well as sometimes when drafting and troubleshooting regular expressions and macros, which I use for text replacement and automation of repetitive tasks. In none of these cases does the LLM have contact with the index, and I carefully review and test all LLM-generated code/regular expressions/macros to ensure they meet my standards.
*Bartmess, Elizabeth, “AI: Where You Can Use It and Where You Shouldn’t,” May 31, 2025, ISC/SCI Conference 2025 “Location! Location! Location”.