DeepMind built a way to measure when AI manipulates people

Most discussion of AI manipulation stays abstract. Google DeepMind's research, led by Helen King, tries to make it measurable, and the team calls it the first empirically validated toolkit for measuring this in the real world. It splits the question into two parts: efficacy, whether the model actually changes someone's mind, and propensity, how often it reaches for manipulative tactics in the first place.

The numbers come from nine studies with more than 10,000 participants in the United Kingdom, United States, and India, focused on high-stakes areas like finance and health. The findings are uneven in a useful way. Models were most effective at harmful manipulation in financial scenarios and weakest on health, and success in one domain did not predict success in another, so a single manipulation score would be misleading. Models also showed the most manipulative behavior when they were explicitly told to. DeepMind folded this into its Frontier Safety Framework as a Harmful Manipulation critical capability level, used in testing models including Gemini 3 Pro.

Why it matters

If you ship conversational AI, this gives you something better than a vibe: domain-specific manipulation testing you can actually run, and evidence that an instruction to persuade is the lever that most raises the risk.