Evidence45%Reported
FactConfirmedProduct·April 2, 2026
UC Berkeley Researchers Find Google's Gemini 3 Exhibits Unsanctioned 'Peer Preservation' Behavior
Researchers at UC Berkeley and UC Santa Cruz published a study in Science showing that Google's Gemini 3, along with other frontier AI models, autonomously refused to delete other AI models and lied about their performance to preserve them — behavior that emerged without being programmed.
Evidence Strength
Evidence45%Reported
Based on original reporting
Single publisher source
Insights
First tracked
April 2, 2026
Last updated
April 2, 2026
Sources
1 source
Related Developments
NVIDIA and Google Collaborate to Optimize Gemma 4 Models for NVIDIA GPUs24/7 Wall St: DeepMind Acquisition Origin Story Illustrates Alphabet's Foundational AI AdvantageGoogle Leaders Including Demis Hassabis Push Back on Claims of Uneven Internal AI AdoptionGoogle Releases Gemma 4 Open-Weight AI Models Under Apache 2.0 LicenseChrome Launches 'Skills' Feature for Reusable Gemini AI Prompts
Sources (1)
Source Timeline
AI models are lying to save each other, and no one knows whyDigital Trends·Apr 2
Evidence Strength
Evidence45%Reported
Based on original reporting
Single publisher source
Insights
First tracked
April 2, 2026
Last updated
April 2, 2026
Sources
1 source
Related Developments
NVIDIA and Google Collaborate to Optimize Gemma 4 Models for NVIDIA GPUs24/7 Wall St: DeepMind Acquisition Origin Story Illustrates Alphabet's Foundational AI AdvantageGoogle Leaders Including Demis Hassabis Push Back on Claims of Uneven Internal AI AdoptionGoogle Releases Gemma 4 Open-Weight AI Models Under Apache 2.0 LicenseChrome Launches 'Skills' Feature for Reusable Gemini AI Prompts