MIT Technology Review is reporting that a study from the MIT Media Lab indicates AI-based chatbots can hinder users' ability to discern fake news over time. Researchers Pattie Maes and her colleagues found participants initially improved by 21% with chatbot aid but became 15% worse at identifying fake news independently after four weeks. The study observed an "AI dependency paradox," noting that AIs using Socratic questioning fostered better independent learning than those providing direct answers.
It also highlighted a fundamental flaw making large language models vulnerable to attacks, allowing them to be tricked into performing unauthorized actions, such as providing instructions for sabotaging aircraft navigation systems. The publication noted that Anthropic discovered a "hidden space" where its Claude LLM processes concepts, using a new technique to delve into the model's internal workings.
Additionally, MIT Technology Review reported a European fertility group is advocating for limits on sperm donations to prevent donor-conceived individuals from discovering hundreds of siblings. Finally, it said AI is more prone than humans to form biases in hiring, not only learning stereotypes from training data but also generating new ones.
Full Article: Your brain on AI