In this session, we explore the tension between harmfulness and harmlessness in AI, and between free expression and safety. We focus particularly on misinformation, or the act of providing incorrect, inaccurate or decontextualized information, and disinformation, the act of creating and spreading false information with the intent to deceive or mislead.

In this video we explore the reasons and the main risks connected to these "hallucinations", or the phenomenon of LLMs creating outputs that are nonsensical or inaccurate. We will talk about the curse of recursion and model collapse, and we will discuss some possible ways to reduce AI-generated misinformation, including mixture of experts (MoE) approaches. 

Disinformation pre-dates generative AI, so what is new now? First, AI can make the act of creating disinformation content much easier and scalable, especially using toxic AIs. Also, AI is extremely good at misleading recipients and creating customized information to make it even more persuadable. Finally, the so called "liar's dividend" shows that in a world were AI content is so accessible, users tend to doubt everything they see online.