Foundation models mark a turning point in the history of artificial intelligence. Before they emerged, every application required a dedicated model trained on task-specific data. Today, a single base model — GPT-4, Claude, Mistral, Llama — can be adapted to dozens of different tasks: text generation, sentiment analysis, classification, translation, automatic summarization.
The term was formalized in 2021 by Stanford HAI (Human-Centered Artificial Intelligence) to describe these models trained through self-supervised learning on massive, unlabeled corpora. The principle is simple: the model learns to predict what comes next in a text from billions of documents, giving it a general understanding of language that applies across many domains.
In the French legal sector, legaltech solutions such as Doctrine, Predictice, Jimini, GenIA-L and Lexis+AI adapt these foundation models to French law. This adaptation relies on fine-tuning on legal corpora and/or RAG architectures that ground answers in verified sources. The challenge is to turn a general language ability into reliable legal expertise.