AI for Lawyers, the fundamentals.

What is an LLM : how it works and its limits

It all starts here. An LLM does not work like a database: it's a statistical model that predicts the most probable continuation of a text. That distinction changes everything. It explains why ChatGPT, Claude or Gemini can produce elegant legal reasoning built on case law that does not exist.

On the agenda : the machine learning behind generative models, the notion of tokens, the context window, and the mechanical reason behind hallucinations. No useless jargon. Explanations designed for lawyers.

Exercise : understanding an LLM's variability

Theory without practice is worthless.

This exercise puts you face to face with a phenomenon every AI user notices without understanding it : ask an LLM the same question twice and you get two different answers. Why ? What does it imply for a lawyer using AI in a brief ? How do you stabilize the results ? You experiment, you observe, you understand.

Overview of the models available in 2026

ChatGPT was never the only model. And depending on the use case, it's rarely the best.

A map of the LLMs available to legal professionals today : GPT (OpenAI), Claude (Anthropic), Gemini (Google), Mistral (Mistral AI), Copilot (Microsoft), GenIA-L (Lefebvre Dalloz). For each model : the strengths, the weaknesses, the price, and the use cases where it excels in a legal context.

The capsule also covers the critical distinction between consumer offerings (free, with no confidentiality guarantee) and professional offerings (API, Enterprise plans, European hosting). A central point for upholding professional secrecy and GDPR compliance.

Why general-purpose LLMs fail at legal reasoning

The capsule that shifts your perspective.

A general-purpose LLM has been trained on the entire web. It knows a little about everything. But legal reasoning has specific features these models do not naturally master : the structured argument, the distinction between facts and law, reasoning by case-law analogy, the hierarchy of norms.

When you ask ChatGPT to draft submissions, it produces something that looks like law. It isn't always. This capsule explains why, and above all how to work around these limits : structured prompts, injecting legal context, forced reasoning chains, source verification.

The architecture of a professional prompt

Prompt engineering separates the lawyer who "tests" ChatGPT from the one who builds it into their daily workflows.

A professional legal prompt rests on an architecture : role, factual context, precise instruction, output format, ethical constraints. This capsule breaks down each element with examples : case-law research, drafting memos, contract analysis, summarizing documents, preparing client meetings.

You leave with reusable prompt templates for your day-to-day practice.

Managing long context : projects, memory, files

An LLM remembers nothing between two conversations. Its context window is limited. And when you feed it an 80-page legal document, it doesn't "read" it the way you do.

These technical constraints have direct consequences for practice. How do you submit a large file to a model ? How do you structure a project across several sessions ? How do you use persistent memory, attached files, system instructions ? You turn your use of AI from an occasional tool into a structured work assistant.

Confidentiality, professional secrecy and GDPR

The capsule every lawyer should watch first.

Is the data you enter into an LLM used to train the model ? Where is it stored ? For how long ? ChatGPT Free, Claude Free and Gemini Free do not offer the same guarantees as professional offerings. And not all professional offerings are equal.

The privacy policies of the main providers, the contractual guarantees to demand, the GDPR framework, sub-processing clauses, data location. The capsule incorporates the CNB's AI tool selection guide (June 2025) and the AI clause in standard fee agreements (April 2026).

The settings that change output quality

Most lawyers use AI with the default settings. It's like driving a car without ever shifting gears.

Temperature, top-p, maximum length, system instructions, stop sequences : these settings directly influence the quality, creativity and reliability of the answers. A low temperature produces deterministic answers (legal research). A high temperature produces creative answers (strategic brainstorming). The optimal setting depends on the task.

This capsule makes these settings accessible to non-technical users and shows their concrete impact on legal use cases.

Summarizing a large case file : method and practical case

The closing capsule is a complete practical case.

You take a large legal file (an expert report, a bundle of exhibits, a set of contracts) and apply every skill you've acquired : model selection, prompt structuring, long-context management, settings, verification, confidentiality.

The goal : to produce a usable summary, the kind of deliverable you could send to a client or use in an internal memo. While measuring the time saved compared with your usual method.