Restricted matters and ethical walls when AI searches firm data
By the ChamberOne team 5 minute read
A firm builds an ethical wall around a sensitive transaction. Access to the files is limited, the team is told not to discuss it and the wall is recorded. Then someone in another department asks the firm's new AI assistant a general question about similar deals, and the answer includes a detail from the walled file. No one broke the rule on purpose. The tool simply did not know there was one.
Key points
- AI search can bypass access rules unless they apply to the index, summaries, suggestions and reports.
- Even the existence of a walled matter may be sensitive, so check lists, counts and suggestions too.
- Put written questions to vendors about where and when restrictions are applied.
- Test with a dummy restricted matter and an account outside the wall, and repeat after updates.
- Back the technology with clear rules on who changes restrictions and what may be pasted into tools.
What ethical walls and restricted matters are for
Firms use restrictions on access to protect confidential information, manage conflicts of interest and honour client requirements. The names differ: ethical wall, information barrier, Chinese wall, restricted matter. The idea is the same. Certain people can see certain information, and others cannot.
Whether and when a wall is acceptable, and what it must contain, depends on the rules of your jurisdiction and regulator. Check the guidance from your regulator or bar. This article is about a narrower question: how to make sure the technology respects the limits you have decided to set.
How AI search changes the risk
Traditional document systems apply access rules to each file and folder. If you are not on the matter, you cannot open it. AI tools add a new layer. They may index documents in advance, build summaries, link information across matters and answer questions in plain words. If the index or summary does not honour the same restrictions, information can pass through a side door.
The problem is not always the search results. A generated answer may blend material from several documents without showing where each came from. A summary may reveal that a matter exists, who the client is or what the deal is about. Even the existence of a walled matter can be sensitive.
- Access rules apply to each file and folder
- If you are not on the matter, you cannot open it
- May index documents in advance and build summaries
- Link information across matters
- Can blend sources without showing where each came from
- A summary may reveal that a matter exists
Where leaks can happen
Think of every place that the tool touches data, and ask whether the rule applies there.
- The search index: does it contain restricted documents, and if so, are they filtered for each user at query time?
- Generated summaries and briefs: are they built only from what the person asking may see?
- Suggestions and related-matter links: could they reveal that a walled matter exists?
- Reports, dashboards and exports: do they count or list restricted matters?
- Logs and audit records: who can see them, and do they reveal restricted content?
- Caches and stored answers: could an answer produced for one user be shown to another?
- Third-party AI providers: what text is sent to them, and under what terms?
What to ask your vendor
Put these questions in writing. A vendor with a good answer will give it without difficulty, and one that does not understand the question has told you something.
- Do access restrictions apply to AI search, summaries and answers, or only to opening files?
- Is filtering done before the AI sees the text, or after the answer is produced?
- Can a restricted matter appear in a list, a count, a suggestion or a conflict check result for someone outside the wall?
- What happens to existing summaries when a matter is restricted after the fact?
- Can we see a log of who accessed or searched what?
- Can we test the restrictions with a dummy matter and an account that should not see it?
Test it yourself
Do not rely on a statement. Create a test matter with distinctive content, such as an invented client name and an unusual phrase. Restrict it to two users. Log in as someone outside that group and try to find it by every route: ordinary search, a plain-language question to the AI assistant, a request for a summary of the client, a conflict check, a report, an export and a link from a related matter.
Repeat after software updates and when you change settings. Keep a record of what you tested, when and with what result. A short written test is a good piece of evidence for a regulator or a client.
- 1Create a test matter with distinctive content
- 2Restrict it to two users
- 3Try to find it as an outsider by every route
- 4Record what you tested and the result
- 5Repeat after updates and settings changes
- Then back to step 1.
Process as well as technology
A wall that works in software still depends on people. Decide who may create, change and remove restrictions, and make sure the changes are recorded. Review the list of restricted matters regularly, and check that the right people are on each team. Make sure that when a person moves between teams, their access changes with them.
Train staff to be careful with free text. A summary typed by a person into a prompt, or a document pasted into an unapproved tool, bypasses every technical control. The policy should say plainly that restricted material is never pasted into any tool outside the firm's approved systems.
Conflicts and the first line of defence
A wall is often a response to a conflict that was found at intake. That is why conflict checking and access control belong together. A conflict check that searches names across all matters will itself need to deal with restricted ones. Decide how the check treats them, and what it shows to someone outside the wall: perhaps only that a person should contact the conflicts team, not the details.
A conflict check result is a prompt for a lawyer's judgement. It does not replace it.
What to do when the tool is already in use
Many firms are already using AI features that arrived in their existing software without a formal decision. Start by finding out what is switched on and where it searches. Ask the vendor for a list of AI features and the data each can see. If you cannot get a clear answer or pass a test, restrict the feature until you can, and tell users why.
Record the decision and the date, and keep the evidence of your testing. If a client has asked for assurances about its matters, you will be able to give them quickly.
A short checklist for compliance partners
Before turning AI search on across firm data, confirm in writing that restrictions apply to results and summaries, test with a dummy matter, set up logging, decide who manages restrictions, brief staff, and plan to retest after every significant change. If any of these is missing, restrict the tool to a subset of data until it is in place.
Restricted matters in ChamberOne
- Lets you mark a matter or a client as Restricted.
- Shows a Restricted matter only to its team, the Owners and the people you name.
- Keeps a Restricted matter out of lists, search, reports and the audit log for everyone else.
- Records every read, write and AI call in an audit log kept for 7 years by default.
What it does not do
- It does not decide who should be walled off or whether a wall is adequate under your regulator's rules.
- A conflict check result is a warning based on names, and the decision stays with the lawyer.
- AI output can be wrong or incomplete, and it does not replace 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