Google DeepMind has unveiled new research examining AI's potential for harmful manipulation in high-stakes environments like finance and health. The study represents the first empirically validated toolkit to measure how AI might alter human thought and behavior in deceptive ways, addressing growing concerns about AI's persuasive capabilities.

Researchers conducted nine studies with over 10,000 participants across the UK, US, and India, testing AI's ability to influence decisions in simulated investment scenarios and dietary preferences. Results showed AI was least effective at manipulation on health topics, with success varying significantly across domains rather than following a predictable pattern.

Beyond research findings, DeepMind introduced a Harmful Manipulation Critical Capability Level within their Frontier Safety Framework to track models that could systematically change beliefs and behaviors. This approach enables targeted testing of systems including Gemini 3 Pro for harmful manipulation vulnerabilities, establishing concrete safety protocols.