AI and legal professional privilege: where the risk sits
By the ChamberOne team 4 minute read
Privilege protects confidential communications between lawyer and client made for the purpose of legal advice or litigation. It is easy to lose by carelessness and hard to regain. So when a lawyer pastes a client's email into an AI tool, a reasonable question is whether that act damages the protection. The honest answer is that it depends, on the law where you practise, on the provider's terms, on how the system is built and on what you sent. This article sets out where the risk sits and what sensible precautions look like. It is general information, not legal advice.
Key points
- Privilege depends on confidentiality, and how AI use affects it varies by country and is still developing.
- Read the provider's real terms on staff access, retention and training before sending anything privileged.
- Send less: strip identifiers and avoid advice-bearing documents and litigation material.
- Approve tools, set data rules, keep a usage log and take advice on high-stakes matters.
The basic worry
Privilege generally depends on confidentiality. If a privileged document is disclosed to a third party in a way that is inconsistent with that confidentiality, the protection may be lost. The question is whether sending text to an AI provider is such a disclosure.
There is no single answer across six countries, and the law on this is still developing. Courts have not settled every point, and the rules differ between legal systems. Check the guidance from your regulator or bar, and take advice on matters where the stakes are high.
Why the provider's terms matter
Think of an AI provider as a service supplier handling your information. The relevant questions are the ones you would ask of a cloud storage provider or an outsourced typing service. Does the provider have a duty of confidentiality? Can its staff read your inputs? Does it keep them, and for how long? Does it use them to train models that other customers use?
A consumer chatbot with terms that allow reuse of inputs is a much worse place for privileged text than an enterprise service with contractual commitments to keep inputs confidential and not to train on them. Read the actual terms. Marketing pages do not count.
- Terms may allow reuse of inputs
- A much worse place for privileged text
- Contractual commitments to keep inputs confidential
- Commitment not to train on inputs
- Read the actual terms, not the marketing pages
Architecture counts too
Terms say what a provider may do. Architecture says what it can do. Where does the text travel? Is it stored? Is it logged by the software between you and the model? Can your firm delete it? A tool that sends only the text needed for a request, and nothing else, has a smaller footprint than one that uploads whole files to a general store.
Ask whether the tool lets you switch AI features off for certain matters, and whether you can see afterwards what was sent. Without that record, you cannot show a court or a client what happened.
What you send matters most
The most dependable precaution is to send less. Strip out names, identifiers and detail that the task does not need. Ask a general question about a clause without including the client's whole contract. Summarise the facts yourself before asking a tool to reason about them.
Be especially careful with documents that record advice, such as an email chain with counsel, and with material prepared for litigation. Treat those as the most sensitive category and do not put them into tools you have not approved.
- Strip out names, identifiers and detail the task does not need
- Ask a general question about a clause
- Summarise the facts yourself first
- Include the client's whole contract
- Put advice email chains or litigation material into unapproved tools
Precautions a firm can take
- Approve a short list of tools after reading their terms, and ban the rest for client material.
- Keep a record of what each approved tool does with data, as reviewed on a given date.
- Set rules on what may be entered: nothing privileged into consumer tools.
- Switch off AI features inside your software where you have not reviewed the terms.
- Keep a usage log so you can show what was processed if a question arises.
- Train staff, using examples from your own matters, on what not to paste.
- Review client and counterparty requirements about AI, especially for in-house teams.
The other direction: AI outputs and your own conversations
Think also about what you create. A client who uses a public chatbot to discuss their own legal problem, and shares the transcript with others, may create records that are not privileged at all. Warn clients about this when it is relevant, especially in disputes where their communications may later be disclosed.
Internal use raises its own question: whether prompts and outputs created in a tool are themselves protected as part of the lawyer's work. That depends on the law and the facts, and it is another reason to keep control of where they are stored.
Questions to put to a provider before you decide
- Do your staff have access to our inputs, and under what conditions?
- How long do you keep inputs and outputs, and can we require deletion?
- Do you use our inputs to train or improve models, for us or for others?
- Where is our text processed, and which sub-processors touch it?
- What do you do if you receive a legal demand for our data?
- Will you tell us before your terms or sub-processors change?
A short scenario
Suppose an in-house counsel uses a free consumer chatbot to tidy an email to an outside lawyer containing the company's candid view of a dispute. The terms allow the provider to keep and review inputs. If that email is later the subject of a disclosure fight, the other side will ask who else saw it. The counsel will have to explain, and the answer will not be comfortable. The same task done in an approved tool with confidentiality commitments, or done by hand, would have raised no such question.
AI controls in ChamberOne
- AI features can be switched off for the firm. With AI features off, nothing is sent to an AI provider.
- Every AI call is recorded in the audit log with the model and token counts.
- When AI features are on, the text needed for a request is sent to the AI provider for that feature, and the Security and AI page says what is sent and on what terms.
What it does not do
- ChamberOne does not give legal advice and cannot tell you whether privilege is affected in your jurisdiction.
- With AI features on, text is sent to OpenAI or Anthropic, outside the country where ChamberOne is hosted.
- AI output can be wrong or incomplete and needs a lawyer's review.
This article is general information about the law, legal practice and technology, current when it was published. It is not legal advice. Laws and guidance change: check the sources linked above and take advice on your own circumstances from a lawyer qualified where you are.
Read next
Client confidentiality and AI tools: a checklist before you paste anything in
The questions to answer about retention, training use, access, hosting and sub-processors before client information goes into any AI tool.
5 minute readShadow AI in law firms: staff are already using it, so what next?
Why bans fail, how to find out what is being used, and how to offer an approved alternative with clear rules.
5 minute readWriting a law firm AI policy: an outline you can adapt
Six sections that a short, usable AI policy needs, with the questions each one should answer.
4 minute read