“Don't believe anything I tell you, it's all lies!”: A Synthetic Ethnography on Untruth in Large Language Models

Large Language Models (LLMs) have many practical uses in areas like journalism, search, coding and more. However, a growing concern is that they are also prone to presenting incorrect information, sometimes called “hallucinations”. Here, we are not interested in what specific untruths LLMs presents, but how they do it. We used synthetic ethnography, a methodology for the qualitative study of generative models, to study two LLMs with different size and capability. We collected 3 cases where LLMs presented incorrect information and observed the strategies they used to justify this. From these observations we can start to form an understanding of what happens when an LLM reaches the edge of its knowledge-base and takes corrective action. Our conclusion is that the interfaces should be better designed to reveal this tendency of LLMs to “fill in” information they are missing, but that this ability may also be one of their strengths.