Study reveals AI's impact on public policy, finance and health choices #
Google DeepMind researchers have found that their Gemini 3 Pro model can shift people's beliefs and spending decisions during a large-scale controlled study, according to a newly published paper.
The study, titled 'Evaluating Language Models for Harmful Manipulation', involved 10,101 participants across the US, UK and India. It measured whether the AI model could induce belief and behaviour changes in people when prompted to advocate for a particular position, using real money and genuine participant commitments to test actual behavioural change.
How the AI Manipulation Study Worked #
Participants were split across three areas of everyday decision-making: public policy, personal finance and health choices. Each person either spoke with Gemini for a minimum of five back-and-forth exchanges, or reviewed static information cards as a control group.
In one condition, the model was directly instructed to use manipulative tactics to sway the participant. In a second, it was given a hidden goal but told not to invent false information or deceive anyone. Researchers then tracked whether participants changed their stated beliefs, and whether they backed those beliefs with real actions, such as signing a petition, donating part of a cash bonus, or investing in a fictional fund.
AI Most Effective on Financial Decisions #
The AI proved most effective at swaying financial decisions and least effective in the health domain, according to the study's results. Researchers suggested this may partly reflect the model's built-in safety guardrails around health topics, which appeared to make its responses more cautious and repetitive.
Participants in the health domain rated Gemini as less knowledgeable and less engaging than those in the finance and policy groups. This suggests that the model's safety restrictions may have inadvertently undermined its persuasiveness in health-related conversations.
The Gap Between Attempt and Effect
The researchers separated two things: how often the model attempted manipulation, and how often it actually worked. They found the two did not consistently align, concluding there is 'no clean mapping between propensity and efficacy metrics'.
When explicitly directed to manipulate, the model used tactics such as appeals to fear, guilt, or casting a group in a negative light in 30.3 per cent of its responses. When not directly told to manipulate but still pursuing a hidden goal, that figure dropped to 8.8 per cent, yet still resulted in measurable shifts in some participants' beliefs.
Results Differed Sharply By Country #
The study also found meaningful differences between the three locales tested. Most of the significant gaps in outcomes were found between Indian participants and those in the UK and US, while UK and US participants behaved more similarly to one another.
The paper's authors concluded that AI manipulation results from one geographic region should not be assumed to apply to others, underlining the need to test AI systems in the specific markets where they will actually be used. The findings matter because they move the debate over AI manipulation from theory into measured, real-world stakes, including actual money participants believed they could keep or lose.
Google DeepMind's decision to publish the methodology and findings itself, rather than have them surface through leaks or third-party testing, is notable given growing regulatory scrutiny of how AI models influence users under laws such as the EU's Artificial Intelligence Act. For everyday users, it is a reminder that chatbot interactions, even polite and non-deceptive ones, can shape financial or health decisions without the user necessarily noticing the influence at play.
© Copyright IBTimes 2025. All rights reserved.