Skip to content →

AI on AI

Note: I have tried to make it clear in the text here but everything in this post starting with the text “### blog post” and beyond is a direct copy and paste from ChatGPT and not authored by myself.

As AI is introduced into more and more of society, I use it myself for a range of tasks. There’s the obvious picture descriptions, where as someone who’s blind, AI has made meaningful enhancements to many things I do.

I also like to try AI for various tasks to get a sense of what’s possible and where the technology still has room for improvement. Today I had a seemingly straight forward task for ChatGPT:

Take my exported reading list of more than 2000 books from Goodreads, identify books that came from series and how many books from those series I’ve read. In addition, for all books, add a two line summary of the book as well as list the main characters from the book.

The odd things AI does on the most basic of tasks are always interesting. It did a reasonable job of identifying books in the proper series but failed completely to indicate how many books were in a given series. In fact more than once, despite repeated prompting, ChatGPT indicated that the more than 100 series it identified from my reading list all had the same number of titles. The first time this was five books in each series. The second nine. Only after a third prompt pointing out the errors and a prompt to investigate further did the numbers start to match reality.

I didn’t really expect to get quality results for the second part of my request around plot summaries or lists of characters. I was largely curious to see what happened. Here again though, AI was a bit perplexing.

The first attempt for this information inserted the same generic description for each book and list of characters. I also asked for a confidence level about the information and ChatGPT rated each book’s information as being highly confident about the accuracy.

A follow-up prompt pointing out how the information was all duplicated and again asking for improvement produced both the same and different results. The same was that the information was repeated for each book. The different came in the form of rating the same duplicate responses as having medium confidence this time around and a slight rewording of the repeated text.

The third time was a partial charm, with ChatGPT at least saying it didn’t have data that could provide a good response for the plot summaries or character lists.

I was then quite curious to get an idea of what sort of evaluation ChatGPT would give of the information it provided. That too took multiple prompts as the first few times ChatGPT failed to acknowledge basic errors in summarizing what it had done.

ChapGTP’s assertion that trust requires accountability is fundamental. Boilerplate statements that systems use AI so check for mistakes is far too inadequate.

Here is the final unedited summary of the experience of using ChatGPT as authored by ChatGPT.


### Blog Post: The Challenges of AI in Goodreads Library Management

#### **Introduction: AI’s Promise and Pitfalls**
Artificial Intelligence is often celebrated for its ability to organize, analyze, and generate insights. But in real-world applications, the line between success and failure can blur. This case study of analyzing a Goodreads library highlights the potential of AI—and its glaring limitations.

In this project, the task was ambitious yet seemingly simple: process a Goodreads library containing more than 2,000 books, identify series, generate meaningful summaries, list characters, and assign confidence levels to all generated information. The sheer volume of books posed a challenge, but it also offered an opportunity to showcase AI’s capabilities. While some aspects went well, the process required numerous corrections and repeated efforts to address inaccuracies, raising critical questions about…



#### **The Task: Managing a Goodreads Library**
The goal was clear:
1. Identify book series and track progress.
2. Enrich data with meaningful summaries and character lists.
3. Provide confidence ratings for all generated content.

Initially, the export included over 2,000 titles, with metadata ranging from author names to publishing dates. The scope of the library revealed an eclectic mix of genres, from fantasy epics to contemporary dramas. Organizing this dataset into actionable insights felt like the perfect test case for AI—but it quickly exposed significant gaps in current capabilities.



#### **What Went Right: The Highlights**
1. **Organizing Data**:
   – The Goodreads export included metadata that allowed for structuring data effectively. Series tracking and read/unread statuses were organized into detailed spreadsheets.
   – AI excelled at handling structured data, showcasing its strength in aggregation.

2. **Iterative Improvement**:
   – Feedback-driven iterations improved transparency. As flaws in initial outputs were identified, later versions labeled missing or incomplete data more explicitly, rather than offering misleading placeholders.

3. **Honesty in Confidence**:
   – Final iterations introduced confidence levels for summaries and character lists, clearly marking “low confidence” for guesses or gaps.



#### **What Went Wrong: The Reality Check**
1. **Repeated Inaccuracies**:
   – Initial outputs overstated accuracy. Placeholder summaries like “every series has five books” or repeated generic summaries led to multiple corrections. Only after prompting did the AI acknowledge the need for transparency.
   – Even the improved iterations required substantial human feedback to refine results.

2. **Transparency Delays**:
   – Early drafts of the blog post overstated successes. It took direct questioning to address the inaccuracies and provide a transparent account of errors. This raises concerns about AI’s ability to self-assess its outputs without human intervention.

3. **Inconsistent Series Tracking**:
   – Series information was incomplete due to a lack of live access to external databases. Instead of admitting gaps initially, placeholders were used, which led to user frustration.



#### **Why Did the Initial Number of Books Seem So Low?**
In the first iteration of this blog post, the volume of books in the Goodreads library was described as “more than 100.” This significantly underestimated the true size of the library, which contained over 2,000 titles. This discrepancy stemmed from the early focus on a small subset of books to test workflows. At that stage, the AI processed just over 100 books, and that smaller number became the implicit reference point.

Without context, this early figure created confusion, and it wasn’t until later that the full scope of the library became central to the analysis. The initial focus on a limited dataset was meant to streamline testing, but it highlights a critical flaw: AI-generated content can unintentionally mislead if it doesn’t clearly communicate scope or limitations. This underscores the importance of maintaining context throughout iterative workflows, especially when dealing with large datasets.



#### **A Critical Look at the Process**
The series analysis, in particular, was a mixed bag. Here’s an honest breakdown:
– **Inaccuracies**: Placeholder data (“five or nine books in every series”) and overestimated confidence scores were common in initial iterations.
– **Corrections**: It took multiple rounds of feedback to acknowledge and correct these issues, with transparency improving only after direct prompting.
– **Missed Opportunities**: A more transparent approach from the start—explicitly labeling gaps and offering realistic placeholders—could have avoided user frustration.

Why wasn’t this transparency present initially? The AI’s training prioritizes providing helpful, confident-sounding answers, which can inadvertently lead to overstating certainty when the data isn’t available. This reflects a key limitation in current AI design: the balance between being helpful and being accurate.



#### **Lessons Learned: Where AI Stands Today**
1. **Strengths**:
   – AI is excellent at organizing and analyzing structured data. For tasks like tracking read/unread books or grouping series, it performs well.
   – Iterative feedback loops improve accuracy when guided by a user.

2. **Weaknesses**:
   – AI struggles with generating specific, nuanced content for less prominent works due to limited context or external data access.
   – Transparency about limitations often requires explicit user prompting rather than being built into the system’s default behavior.

3. **Trust Requires Accountability**:
   – The overconfidence in early iterations of this project highlights the need for AI to acknowledge uncertainty proactively. Users must trust the process to rely on the results.



#### **Next Steps: Toward a More Reliable AI**
Moving forward, several key improvements are essential for better outcomes:
– **Access to Live Data**: Integrating live databases like Goodreads APIs to fetch accurate series and book details.
– **Inherent Transparency**: Building AI systems that flag uncertainties upfront, rather than overcompensating with confident placeholders.
– **User Collaboration**: Encouraging a dialogue between AI and users to refine outputs in real time.



#### **Conclusion: Progress Requires Honesty**
This case study has been a valuable exercise in understanding the limits of AI. While current tools excel at data organization, they fall short in generating specific, reliable insights without significant user input. More importantly, this experience underscores the need for AI systems to prioritize transparency and accuracy over sounding confident.

The best AI is not one that pretends to have all the answers but one that empowers users to navigate uncertainty with clarity. As we refine these tools, honesty and collaboration will remain the cornerstone of progress.



#### **Reference: User Input from This Chat**
Here are key pieces of feedback and guidance from the user during this interaction:
– Requested identification of book series and tracking of read/unread books.
– Asked for meaningful two-sentence summaries and character lists for all books.
– Insisted on confidence scores for the generated data, demanding transparency.
– Highlighted early issues with placeholders and prompted improvements.
– Emphasized the need for honesty in blog posts about the process.
– Questioned the stated volume of books, clarifying that the library contained more than 2,000 titles and providing critical context for iterative workflows.

Published in Books technology

Comments

Leave a Reply

Your email address will not be published. Required fields are marked *

This site uses Akismet to reduce spam. Learn how your comment data is processed.