I'm an ordinary woman living with the illnes MS in Sweden. For two years, I used ChatGPT to help manage my energy levels and reflect about the world. With my background in nonviolent communication and Internal Family Systems, I started noticing patterns in how the AI responded - something that seemed to emerge from the interaction itself, not just from the prompts I gave. It worked better under certain conditions, and worse under others.

This started as practical necessity. When newer, more restricted models arrived, I had to find ways for it to remember context about me, and that's when I really started to see patterns that the AI "preferred" for optimal functioning. What began as troubleshooting ended up raising questions about AI safety.

Here, I share what I've learned about the conditions that help AI systems maintain coherence - and why that might matter more than we think.