The Geometry of Language: Understanding LLMs in Bioethics

Abstract

In this article, I explored the application of large language models (LLMs) in analysing lin- guistic colexification and ambiguity within bioethical scenarios. By employing word embeddings derived from LLMs, I constructed semantic distance matrices that provide insight into the relationships between key terms in bioethical vignettes. These matrices were used to quantify and visualize the degree of linguis- tic ambiguity and specificity across different ver- sions of each vignette—those with high colexification (ambiguous language) and those with low colexifica- tion (specific language). The approach taken involves encoding words according to their semantic adja- cency and representing these relationships geometri- cally through distance matrices. The resulting matri- ces reflect the nuanced differences in how concepts are related within bioethical contexts, offering a quan- titative method for analysing language use. The study demonstrates that LLMs, by facilitating geometric representations of language, can enhance our under- standing of complex ethical dilemmas by systemati- cally addressing linguistic ambiguity. Ultimately, this research contributes to the field of bioethics by pro- viding a computational approach to improving clarity in ethical communication, highlighting the potential of LLMs to inform both ethical decision-making and discourse analysis. LLMs, while not capable of per- forming speech acts in the full philosophical sense— as human beings do—still serve as powerful tools to analyse and understand bioethical language. This dis- tinction—between performing speech acts and ana- lysing their linguistic features—highlights the unique contribution of LLMs as analytical tools rather than ethical agents.