USING AI FOR LEGAL RESEARCH
Introduction
Training vs conversation
Uploading documents
Using AI is no substitute for obtaining proper advice and help from a qualified lawyer. However, as I often work directly with clients and other professionals, I thought it useful to explain how to use AI as a tool. It is now common for parties, including lawyers, to use AI for research and drafting.
This is not a comprehensive explanation. I asked ChatGPT (the subscription version) about how it treats information, especially uploaded documents.
Like many lawyers, I use Lexis-Nexis' Protégé, a subscription legal research service which public AIs cannot access.
Uploading documents
Training vs conversation
Uploading documents
A document is not published on the internet when it is uploaded, or contents pasted into a question.
- It is not indexed by search engines
- It is not made available to other ChatGPT users
- It does not become part of any public dataset
Another user cannot retrieve the uploaded document by asking similar questions.
Training vs conversation
Training vs conversation
Training vs conversation
a) Within a chat, the AI:
- can use what is provided in the conversation to answer
- does not have a persistent personal memory unless explicitly enabled and appropriate
b) Training/improvement
- OpenAI may use conversations to improve models unless the user is using an enterprise/business tier with training disabled
- Even then, the model does not store or reproduce documents verbatim in normal operation
AI memory
Training vs conversation
Training vs conversation
Critically the AI:
- does not have access to other users’ chats
- does not recall past specific documents
- cannot intentionally reproduce a document someone else uploaded
USING AI FOR LEGAL RESEARCH
Verbatim leakage
Verbatim leakage
Verbatim leakage
For professional risk assessment:
- The system is designed not to emit memorised user documents
- The practical risk of an exact document appearing to another user is extremely low
- However, for highly sensitive or privileged material, best practice in the legal profession is still to treat any third-party AI as a confidential processor, not as legally privileged storage
Good practice
Verbatim leakage
Verbatim leakage
Barristers and solicitors commonly:
- redact client identifiers
- avoid including unnecessary personal data
- use enterprise/private AI tiers where available
- treat outputs as draft assistance, not filing-ready text
Summary
Verbatim leakage
Conclusion
Information used to ask an AI questions, or documents uploaded for AI comment, are:
- ❌ Not public
- ❌ Not searchable by others
- ❌ Not shared between users
- ⚠️ Still involves a third-party processor, so apply normal confidentiality judgment
- ✅ Generally acceptable for anonymised drafting
Conclusion
Verbatim leakage
Conclusion
There is nothing wrong with using modern tools to help analyse and prepare a case. Nonetheless, professionals do not substitute AI output for their judgment, which is informed by training and experience.
Be careful how AI is questioned or tasked to avoid inputting biased or incorrect information because a particular answer is sought. Seek a neutral evaluation or assessment to understand strengths as well as weaknesses. Remember, garbage in, garbage out creates unreliable output.
There remains a crucial role for professional input, even if assisted by a computer.