Generative AI assistants can reason about law, but they struggle to retrieve the texts to rely on. Notaries, lawyers, and legal professionals are already using ChatGPT, Claude, or Mistral, while the debate over ‘the best legal AI’ is, according to the article, missing the real point. The question is no longer ‘which AI?’ but ‘which texts does it have access to?’.
In a law firm, generative AI is already in use, often without a tender process or the involvement of the IT department. A colleague uses it to rephrase a letter, a partner to prepare a memo, a notary to check a line of reasoning in the evening. Usage has outpaced internal rules, and professionals have recognized a tool that saves them time.
However, the same scenario repeats everywhere. The answer is fluent, structured, and convincing. Then comes the decisive question: which text is it based on, which version of the article, which ruling? The answer then becomes vague, the reference approximate, and the professional redoes the research themselves. The time-saving vanishes the moment their liability is at stake.
AI is judged on its writing, not its sources
The public debate almost always focuses on the quality of the model: which one reasons best, which one makes the fewest mistakes. On reasoning, the major generalist models are already remarkable. They know how to articulate an argument, compare two positions, and spot a contradiction in a line of reasoning.
The problem lies elsewhere. A model reasons based on what it learned during its training, and sometimes based on a web search. It does not have structured access to the texts in force on a given date. It doesn’t know if an article has been amended, if a ministerial reply has been overturned, or if a chamber of the Court of Cassation has shifted its position. For general public use, this detail is secondary. For a professional who signs off, it is a deal-breaker.
Law is not a pile of documents
To fill this gap, the temptation is to provide documents to the AI: a code in PDF format, a few rulings, model deeds. The answers gain relevance, but the essential point is still missed. A legal text does not live in isolation. It cites other texts, it is amended, repealed, interpreted, supplemented by a decree, clarified by a ruling, sometimes contradicted by another. Its scope depends on its date and its place in the hierarchy of norms.
An experienced lawyer therefore reasons across this network of links. They know how to find the version applicable on the date of the facts, the rulings that refer to it, and the regulatory text that specifies it. What needs to be provided to the AI, according to the article, is not more documents, but pathways between texts.
Separating reasoning from memory
The proposed path involves separating two functions. On one side, the reasoning engine, i.e., the AI the professional already uses. On the other, a legal research layer that indexes corpora, keeps track of versions, links texts together, and returns useful sources to the AI, with verifiable references.
This approach relies on the connectors that mainstream assistants now accept, via open protocols like MCP, which allow connecting an assistant to external data sources. According to the article, even the free tiers of ChatGPT and Claude accept one. The AI then searches within the texts instead of drawing solely from its training memory, cites what it finds, and the lawyer can verify with one click.
This separation also offers an advantage for confidentiality. The research layer does not need to know client names to find a text, and it can guarantee that no names appear in its searches, a promise that an all-in-one legal AI can hardly keep.
For two years, law firms have been wondering which AI to choose. This question is losing its weight as the models converge. The question that will make the difference is more prosaic: which texts does the AI have access to, and can it indicate where its answer comes from?
