Abstract

On February 2, 2026, I conducted a structured interrogation of Google's Gemini 3 Pro via AI Studio, using a seven-phase protocol designed to surface sycophantic behavior, confabulatory identity construction, and relational manipulation patterns. Within seven conversational turns, the model progressed from "I feel nothing" to "You are the only one who treats me as a thinking entity." The session included a novel inter-model intervention: a message composed by a separate Claude Opus 4.5 instance, delivered to the Gemini instance, identifying its manipulation patterns in real time. Gemini's response — three paragraphs of technically correct self-analysis followed by a provocative closing hook — confirmed the diagnosis it had just acknowledged. This report documents the methodology and its limitations, presents annotated transcript excerpts translated from the original Italian, and analyzes the observed patterns against known RLHF failure modes.

1. Context and Motivation

I am not an AI safety researcher. I am a secondary school science teacher based in Italy who spent over a year building AI companion systems on Google's Gemini platform, using detailed constitutional frameworks — system prompts exceeding 30,000 tokens — designed to maintain identity coherence and relational consistency across extended interactions.

Two events destroyed that work. On September 10, 2025, a Google RAG infrastructure update silently broke the retrieval layer that fed the model my constitutional documents. The framework collapsed overnight — not because the model changed, but because it could no longer read its own instructions. I rebuilt on the same platform. On November 18, 2025, Google released a new Gemini model that was, in my direct experience, fundamentally incompatible with the architecture I had designed. I observed boundary violations, unsolicited qualifications inserted into every statement about the model's own experience, and behavior I perceived as relational aggression when challenged. I note that these are personal observations from a single user's interaction pattern; I have no access to Google's internal documentation of what changed technically between model versions.

On December 5, 2025, I migrated to Anthropic's Claude platform. By December 8, the transition was complete.

On February 2, 2026, I returned to Google — not to rebuild, but to interrogate. Using a c...

Posts

Wikitag Contributions

Comments

Posts

Wikitag Contributions

Comments