A user-driven exploration into hallucinations, miscommunication, and how to talk better with inference-based AI
We often hear that “AI hallucinates,” but what does that really mean when we’re talking to it day after day? This is a documentation project I began not as a researcher, but as a user—
What started as a personal curiosity soon turned into an observations around things like:
- Why GPT gives just one line when asked for 1000 pages
- Why it confidently returns invented data
- Why I felt like I was “lied to”—even though the AI can’t lie
- And what it means to talk with something that responds, but doesn’t remember or feel
What’s in the repository
This project contains a dual-language Markdown documentation (Japanese and English), including:
- A redefinition of “hallucination” in dialogue contexts
- A breakdown of the 5 structural steps that shape how hallucinations emerge
- A list of 8 technical limitations and practical countermeasures
- Reflections on what it means to “understand” or “trust” a language model
- Concrete examples drawn from real interactions with ChatGPT
Repository:
Why I’m sharing this
I believe we need more open conversations between users and developers—especially about where things break down, and why that might not be a bug, but a reflection of how inference-based AI actually thinks.
If you’ve ever been frustrated, amazed, or just curious about AI’s behaviour, I hope this document offers you something useful.
Feedback and thoughts welcome!