Large Language Model (LLM)

A large language model is a deep-learning model trained on enormous text corpora in order to predict the most probable word, or more precisely token, given the preceding context. From this apparently simple task emerge complex capabilities of writing, translation, summarization, apparent reasoning and code generation. LLMs generally rest on the transformer architecture, which lets them weigh the relative importance of the different elements of a text. Their operation is fundamentally probabilistic, they do not consult a knowledge base but statistically produce plausible sequences, which explains both their fluency and their propensity for confident error, or hallucination. For insurance and risk management, the LLM is the cognitive engine of generative and agentic AI systems, and therefore the source of new risks, opacity of reasoning, behavior unpredictable at large scales, and vulnerability to semantic manipulation. An important nuance must be stated: the LLM is neither a database nor a reliable logical reasoner, and mistaking the fluency of its expression for a guarantee of accuracy is a frequent and consequential error of judgment in critical professional uses.

A firm uses a large language model to pre-analyze insurance contracts; the model produces fluent summaries but occasionally introduces non-existent clauses, which makes systematic human review essential.

LLM, grand modèle de langage, large language model, modèle de fondation