Various players providing online search services, both established and emerging, seem to be betting on large language models that generate information relevant to a user's query as the future of online information retrieval. Additionally, researchers in various fields are being encouraged to use LLMs for various research activities. Given the demonstrable current limitations of such systems, not least of which are their tendency to hallucinate and our tendency to anthropomorphize them, the question arises as to whether general LLM-based online search and research in general is a good idea, especially in the short term. In this paper, I assess the potential risks of LLM-based search and weigh them up against the potential benefits. My main focus is on contexts where users engage in search/research informationally—i.e. where they are looking for accurate, factual information on a topic. I draw on work done in social epistemology on knowledge acquisition and the nature of expertise to argue that LLM-based systems for both online search and more specialised research pose significant epistemic and moral risks. At the same time, the supposed benefits of such use cases are in most instances relatively insignificant. I argue that LLM-based search and research are currently not a good idea.
Time: February 08, 2024, 14:00-15:30
Location: Online (Zoom; registration at jan.hinrichsen@med.uni-goettingen.de)
Lecturer: Dr. Tanya de Villiers-Botha (Stellenbosch University, South Africa)
Organizer: Institute for Ethics and History of Medicine, Göttingen & Institute of Advanced Studies for Ethics of Science
Further information can be found here