School of Computer Science Seminar Alan Davoust - On AI-generated mis- and disinformation

Alan Davoust will present On AI-generated mis- and disinformation.

Alan Davoust is an associate professor of computer science at Université du Québec en Outaouais (UQO), in Gatineau, Canada. He is also a member of the INRS-UQO Joint Research Unit (UMR) in Cybersecurity and Digital trust. After studying software engineering in France, he obtained a master's degree and then a PhD from Carleton University (Ottawa, Canada) in 2015. His thesis focused on peer-to-peer computing and the decentralised control of information. He then spent two years as a postdoctoral researcher at the University of Edinburgh, where he participated in large-scale projects on socio-technical systems and algorithmic bias. His current research focuses on multi-agent and socio-technical systems, ethical aspects of AI as well as AI-produced misinformation and disinformation.

Abstract: We explore three problems related to LLM chatbots and false information.

The first problem is their tendency to hallucinate, i.e. produce incorrect answers to factual questions. Retrieval Augmented Generation (RAG) is generally considered to be a useful mitigation to this problem, but traditional RAG can be affected by noise and contradictions in the external knowledge. Here we show that Graph RAG, a setup where external knowledge is provided via graph-structured data, is much more robust to such noise.

The second problem that we consider is that of LLM-produced disinformation, i.e. intentionally misleading claims or “fake news”. In a context where we attempt to detect such fake claims, we show that LLM-generated disinformation is significantly more difficult to detect via traditional machine learning methods, and discuss possible mitigation methods.

Finally, we consider the problem of ideologically-loaded discourses, a form of influence that goes beyond the truth value of information. We describe an exploratory study in the RAG context, aiming to assess to what extent ideologies present in external knowledge transfer to LLM answers, and whether this can be controlled by prompt engineering.