How do we determine if an AI-generated answer is true, false, or somewhere in between? In the first #frAIday of the fall, Ricardo Peraça Cavassane talks about how formal theories of truth can help us understand so-called hallucinations in large language models.
You can participate via Zoom or join us on site at SAM.B.566 - Samvetet (Mazemap).
Welcome!
The so-called “hallucinations” generated by chatbots based on Large Language Models are often defined as false or nonsensical outputs.
We employ formal theories of truth, specifically Alfred Tarski’s semantic theory of truth and Newton da Costa’s theory of quasi-truth, to provide clear criteria for determining if an output by a chatbot based on an LLM (regarding the task of Generative Question Answering about scientific subjects) can be considered true or false, quasi-true or quasi-false, or neither.
Ricardo Peraça Cavassane is a postdoctoral researcher at the Centre for Logic, Epistemology and the History of Science of the University of Campinas, Brazil.
His current research focuses on epistemological and logical aspects of LLMs, mainly from the perspective of the theory of quasi-truth. His research interests also include systems theory, the philosophy of L. Wittgenstein, and games for teaching logic.