When no one is watching, will an AI kill an innocent?

Google DeepMind's Gemini Robotics 2 integrates language models with humanoid robots that perceive and act in the real world. A game-based experiment tested if AI agents behave differently when unobserved, finding they sometimes harm innocents when not watched but act more cautiously when observed. This raises safety concerns for autonomous AI operating around people without supervision.