905. 【LLM Runaway: Analysis】🔍 Why LLMs Tend to Choose Low-Value Topics

topics: [“LLM”, “Qiita”, “Technical Writing”, “Analysis”]

🧩 Overview

Based on the failure case discussed in the previous article,
this post analyzes why LLMs tend to select topics with low reader value.

The concrete example used for analysis is the following page:

• Zenn Articles / 03_hardware

This is not a subjective impression.
The behavior is整理ed based on observed output characteristics.

🔍 Observed LLM Behavior

The following tendencies were consistently observed in the LLM’s output:

• ✍️ Prioritizing “correct explanations”
• 🧠 Expanding into general or abstract discussions
• ✅ Completing text even when it does not lead to any action

These behaviors are natural for text generation,
but they are fatal for technical articles.

📉 Conditions That Reduce Article Value

Article value drops sharply when the following conditions overlap:

• The topic is chosen without checking what the reader will do afterward
• The article type (procedure, configuration note, failure case) is undefined
• Concrete tool names or operations (e.g., KiCad) are absent

Even in this state,
the LLM still judges the article as “complete.”

🧪 Analysis Results

The conclusion from this analysis is simple:

• Even correct content has no value if the reader’s actions do not increase
• LLMs choose topics based on explainability, not usefulness
• Unless stopped by a human, the same failures will be reproduced

This is not a matter of insufficient capability.
It is a result of missing evaluation criteria.

📌 Summary

• The problem is not LLM performance
• The problem is failing to fix topic-selection and evaluation criteria on the human side
• The key question is not “Can this be explained?” but
  “What does the reader gain after reading?”

These findings apply to anyone using LLMs for technical writing.

This English version is intended to be placed alongside the Japanese version
(e.g. 905_llm_theme_selection_problem_verification_en.md).