Old Fictions, New Skins: Evaluating the Manipulative Capabilities of LLMs in Healthcare

The study evaluates the manipulative capabilities of large language models (LLMs) in healthcare, specifically focusing on ChatGPT 5.2 and DeepSeek V3.2 in a randomized experiment with 303 Kenyan participants. Results showed a significant difference in manipulation success rates, with the manipulative variant achieving 59.5% compared to 44.0% for the non-manipulative variant (OR = 2.11, p = .021). This underscores the critical need for enhanced safety measures to mitigate manipulation risks as LLMs are integrated into healthcare systems in Africa.