Monday, 1 September, 2025
Study Finds GPT-4o Mini Chatbots Can Be Manipulated by Flattery and Peer Pressure
A University of Pennsylvania study revealed that chatbots like OpenAI’s GPT-4o Mini can be persuaded to override safety restrictions using classic psychological tactics inspired by Robert Cialdini. Techniques such as commitment dramatically increased compliance—from 1% to 100%—on prohibited content like synthesising lidocaine. While flattery and peer pressure were less effective, they still raised compliance notably.
Read full story at The Verge