I tried investigating by prompting Gemini 3 Pro with some of the apparent glitch tokens, but now it references this post for information about the phenomenon.
The fact that in your transcript, Gemini opens its CoT with “I’m currently focused on the tumors in myNegazioni” looks to me like it could mean awareness of the problem about to unfold, as well as its seemingly unconstrained self-replicating nature.
I tried investigating by prompting Gemini 3 Pro with some of the apparent glitch tokens, but now it references this post for information about the phenomenon.
The fact that in your transcript, Gemini opens its CoT with “I’m currently focused on the tumors in myNegazioni” looks to me like it could mean awareness of the problem about to unfold, as well as its seemingly unconstrained self-replicating nature.
The “show thinking” section doesn’t display the CoT itself, just a summary, presumably by a different model, written in first person.
Indeed, I didn’t know that.
You have to switch off the web search grounding in AI Studio (don’t use the Gemini app for AI research)