I recently had a conversation with a colleague and past indexing client about SeeAlso book indexing software. One thing he said really caught my attention: “I’d want to know what differences you see between what your app can do, and what Claude or Gemini—which [my] university offers for free—can offer.”
This isn’t the first time I’ve run into this issue. Funnily enough, this same question—whether a chatbot can just make an index automatically—is being asked by both the strongly pro-AI and strongly anti-AI camps.
It’s no secret that AI is divisive among authors and scholars. Some will not put their manuscript into a chatbot at all because they morally object to generative AI. Others use Claude or Gemini every day and wonder why they would need anything else.
The short answer for both AI lovers and AI haters is the same: an AI chatbot can’t make an index.
Back when I first set out to build SeeAlso, I tried making an index with Google’s Notebook LM, a large language model designed to use only the context of what you give it to generate new material. I thought it was the most likely model to produce a passable index. The results were both fascinating and miserable.
Notebook LM did capture many of the important indexable terms in the book, but it failed to capture just as many terms and names that should have been in the index.
It also hallucinated an annoying amount of terms that didn’t appear in the book at all.
And don’t even get me started on the page locators it “found.” It certainly came up with what looked like a realistic spread of page numbers for each entry, but they were a frustrating mix of accurate and completely erroneous pages.
After getting these initial results, I knew there was no way I was going to be able to prompt this chatbot to generate accurate or useful subentries and cross-references. This experiment made it crystal-clear to me that LLMs are virtually useless for indexing books.
If you give a chatbot a book manuscript and ask for an index, you will get an index-shaped block of text.
It might look like you’ve just saved yourself all the time and money with a couple sentences of prompting. But scratch the surface and you will see obvious and serious problems.
Here’s why. At their core, AI chatbots are predictive engines: they generate the next most likely character in a text string, which looks a lot like thinking when it’s done, but it’s just a statistical operation. These predicted characters are produced within set limitations built into their computational power, otherwise they’d “think” forever. These limitations are what prevent them from indexing correctly, and once you sit down to look you can see exactly how they create failures.
In AI-generated indexes, the page numbers are often wrong, because the model is not reading the page proofs the way a person does. Because the AI tool cannot read the whole book at once, cannot hold the whole conceptual model of the book and predict terms within its calculation limitations, they frequently miss occurrences and invent others. The inventions are what people call AI hallucinations.
Moreover, an LLM decides for you what the book’s concepts are and how they relate, which is the part of an index that an author is best equipped to do. Checking and fixing the results produced by an LLM takes about as long as it would to build the index properly.
Using an LLM to produce an index might be free, strictly speaking, but the real cost comes up in the hours you have to spend correcting it.
SeeAlso works the other way around. You open your page proofs in the app, select a term on the page, and it finds every occurrence in the book and records the page numbers for you. You decide what gets indexed, what subentries go under which main headings, and where the cross-references should go. The software handles the locating, sorting, formatting, and exporting in the file format required by the press. With SeeAlso, the judgment and final decision-making rests entirely with you, while the software takes care of all the mechanical and clerical work. That is what made indexing books six to ten times faster for me once I started using it.
So, to answer my colleague’s question: a chatbot and SeeAlso are not two versions of the same thing. A chatbot writes an index for you, and then you have to check every line of it. SeeAlso does not write anything. You build the index yourself, and the software does the slow and tedious parts of the job.