Abstract: 

LLM products are increasingly used to as part of research, even naming one of their features “deep research”.

Searching for journal articles and facts on the web, they engage in structured practices of interpretation, curation, interpretation, curation and interpretation. This talk discusses the many paths through which generative text processes (LLMs) are divergent, resulting in different models, different prompts, or the same model-prompt generating importantly different answers. This is critically important for medical questions, and anything with uncertainty or indeterminacy as part of its foundation (business, science, politics, environment, war). Summaries are not neutral, especially with regard to how they treat uncertainty, especially when iterative prompt responses drift among demands. 

Perhaps we take seriously LLMs as provocative spectres rather than letting our scholarship be abducted by unthinking aliens.