Why GPT-5 Makes “Wrong Maps” and “Gibberish Text” — And Why It’s Not the LLM’s Fault

If you’ve played with ChatGPT’s image generation features, you’ve probably seen it: coastlines that look suspiciously unfamiliar, countries with oddly shifted borders, or text in images that reads like it’s from another alphabet. Many people walk away from that experience thinking: “Wow, GPT got that completely wrong.” But here’s the twist — in most cases, GPT never drew anything in the first place.

For those without a Medium subscription, I’ve shared a friend link: https://0xhagen.medium.com/why-gpt-5-makes-wrong-maps-and-gibberish-text-and-why-its-not-the-llm-s-fault-bfb4b71763de?sk=003564dfca25246145289681b261dfb6

Before I’m going to explain more in detail, let’s draft the workflow:

Internal Workflow and Tooling

The process and all required tools behind image or map creation after you submit a prompt looks roughly like this:

LLMs vs. Image Models: Two Very Different Brains

• LLM (Large Language Model) → Trained on billions of text sequences. It’s great at:
  - Understanding your request
  - Reasoning over information
  - Generating coherent text and instructions
• Diffusion Model (or similar image…