The Language of Large Language Models: Observations, Mechanisms, Wider Implications

Large Language Models have their own ways of using language, and those ways are not always in line with human language usage. Many models overuse em-dashes, they 'emphasise' things, and favour constructions such as 'X, not Y'. This talk explores the linguistic features of LLMs and uses some of these features, such as formatting and word choices, as a tracer dye to help understand model mechanisms: why models exhibit certain behaviours, how models differ from each other, how model behaviour varies across languages, and how models learn and unlearn. Wider implications are explored: are there potential influences on human language and beliefs, and what can be learnt about model learning in general?