Study shows AI can be trained to fake UN political speeches

Researchers from the United Nations Global Pulse conducted a case study demonstrating that artificial intelligence (AI) can be trained to generate remarks in the style of UN political leaders. In their experiment, they used speeches given by political leaders at the UN General Assembly between 1970 and 2015 as training data for an automated speech generation tool. The language model was then successfully used to generate text on general topics (such as climate change) as well as remarks from specific leaders such as the from UN Secretary-General. In these cases, coherent text was generated around 90% of the time and the language model was also able to accurately include contextual information. Based on their experiment, the researchers drew attention to the potential dangers of automated text generation, including in the context of peace and political stability. They also outlined a few recommendations for addressing these dangers, such as mapping the potential human rights implications of these technologies, developing tools for continuously monitoring AI generated content, and setting up strategies for countermeasures and scenario planning for critical situations.​