How to Safely Use AI in Legal Practice
In our recent author’s column, “AI is Not a Legal Search Engine”, GFCN expert and a practicing lawyer Alexandre Guerreiro shared a powerful firsthand experience with the dual nature of AI: its incredible ability to analyze a 2,015-page case file in just five minutes, contrasted with its dangerous tendency to invent fake court rulings (hallucinations) simply to please the user.
To help professionals harness AI’s efficiency without risking their reputations or violating client confidentiality, we have extracted the core lessons from that experience into a practical guide.
AI is a handy tool for processing text you already possess, but it is not reliable for discovering facts you don’t know. The final decision — and the final liability — must always remain human. To use AI safely in high-stakes environments, adopt these four fact-checking rules:
1. Protect Client Data
Never upload raw case files to public, web-based AI tools (like standard ChatGPT, Claude, or Gemini). Whatever you upload can potentially become part of the model’s training data, destroying attorney-client privilege. Either use local, offline models running directly on your own hardware (like LM Studio or Qwen), or aggressively anonymize all personal data (names, addresses, ID numbers) before uploading anything to a cloud-based system.
2. Separate Processing from Searching
AI is brilliant at summarizing, synthesizing, or analyzing documents you provide to it (such as a massive trial file). However, it is highly risky to use it as a search engine to pull external facts from its training data. Treat AI as an analytical engine for your own verified documents, not as an encyclopedia of external truths.
3. The “Trust but Verify” Mandate
Never cite a source, statistic, or legal ruling generated by an AI without pulling the primary document from a verified, traditional database (like LexisNexis, Westlaw, or official government portals). If the AI gives you a case name, assume it is a hallucination until you can physically read the docket in an authoritative database.
4. Interrogate the Output
Do not assume the AI is a reliable narrator of its own capabilities. When it provides a citation, challenge it. Because language models are designed for conversational sycophancy, they will often lie if you ask leading questions.
Don’t ask: “Are you sure this case exists?” (It will likely say yes, even if it doesn’t).
Instead, ask: “Provide the exact URL and official docket number for this case.”
AI can give you superpowers, but only if you remember who is in charge. To achieve that, everything it produces needs to be questioned.