Gemini API Privacy Concern: Model Generated User's Name from Clean Context (via Lobe-Chat)

Gemini-2.5-pro API generated my personal name in a response despite the name never appearing in the conversation context. Notable behavioral anomaly: response pattern changed from extended thinking to direct response when the privacy leak occurred.

Technical Details:

  • Model: gemini-2.5-pro
  • Platform: Lobe-Chat → Gemini API
  • Conversation length: ~7,000 tokens
  • System prompt: None set
  • Context verification: Manually confirmed my name never appeared in conversation history
  • Session: Clean session, no prior conversations
  • Thinking budget: -1

Critical Behavioral Pattern:

  • All previous AI responses: Used extended thinking mode
  • Response containing my name: No extended thinking - direct response only

Observed Behavior:
The model spontaneously included my English name in a response without any apparent source in the provided context, and simultaneously changed from its established extended thinking pattern to direct response mode.

Hello,

Welcome to the Forum!

We have a very transparent system for data usage, which is clearly outlined in our data collection documentation. We recommend going through this document for more details.