Blog
Can AI help us redact documents before they go out?
Yes, for the finding and listing part: AI can flag candidate passages and tell you which page they are on. The judgement about what comes out, and the act of redacting the file itself, stay with a human.
Alesis · · 5 min read
Yes, up to a point. AI is useful for the first half of a redaction job: reading a set of documents, flagging passages that may need to come out, and telling you which page each one is on, so a fee earner is not scanning nine hundred pages cold. It is not safe to leave the second half to a tool. The decision about what should be redacted, and the act of applying the redaction to the file so that the text underneath is genuinely gone, belong to a person who understands the consequences of getting it wrong.
What the AI part of the job is good for
Redaction work splits into finding, deciding and applying. AI helps most with finding.
- Sweeping a long set for categories you have defined: bank and card details, national insurance numbers, dates of birth, home addresses, third party names, medical content, material that appears to record legal advice.
- Consistency across a bundle. A name that appears on page 12 and again on page 480 is the kind of thing a tired reader misses on a Friday afternoon.
- Producing a working schedule: the page, the passage, the category, so the fee earner is making decisions from a list rather than from memory.
- Telling you which pages it could not read properly. Scanned faxes, photographs of letters and handwritten notes are exactly where redaction failures happen, and you want those pages named rather than silently passed over.
None of that is a decision. It is preparation for a decision.
Where the real risk sits
Two failures cause the problems firms actually have.
The first is incomplete finding. A tool that lists thirty candidate passages has not told you there are only thirty. Treat any list as a starting point, not a guarantee of completeness, and read the document yourself if the material is sensitive or the set is small enough to allow it.
The second, and the more damaging, is bad application. Material that has been visually covered is not redacted. Black boxes drawn over text in some document tools leave the text selectable and searchable underneath. The same applies to hidden rows and columns in spreadsheets, tracked changes, comments, document properties and other metadata, and to the original file sitting in a document management system next to the redacted one with a very similar name. These are production problems, not AI problems, and they are solved in your document tools and your process, not by asking a model to do it.
There is also the judgement question. Whether material is properly withheld, and on what basis, is a matter for the fee earner with conduct, taking account of the duties that apply to the disclosure or request in hand. The Civil Procedure Rules, UK GDPR and Information Commissioner's Office guidance all bear on different parts of this, and none of them is satisfied by a tool's opinion.
A process that holds up
- Decide the categories first. Write down what you are redacting and the reason for each category before anyone touches the documents. If you cannot state the reason, you are not ready to redact.
- Run the sweep. Ask for candidate passages with page references and a reason for each, and ask for unreadable pages to be listed separately.
- Review every candidate. The fee earner accepts, rejects or amends each one and adds anything the sweep missed. This is where the time is spent and where it should be.
- Build the schedule. Page, passage, category, reason, decision maker. You will want this if the redactions are later challenged or questioned.
- Apply the redactions in the document software, properly, and flatten or export so the underlying text is removed rather than hidden.
- Test the produced file. Search it for the names and numbers you removed. Open it on a different machine. Check the metadata and the document properties. Check the page count matches.
- Have a second person check a sample, and more than a sample if the material is sensitive. Record who did the work and who signed it off.
Bulk sets and repeat jobs
On a large set, the honest position is that nobody reads every page with equal care, and redaction errors cluster in the places attention drops: the back half of exhibits, appendices, duplicate email chains, spreadsheets. AI helps because it gives the same level of attention to page 700 as to page 7, and because it can tell you where duplicates sit so one decision can be applied consistently.
It does not reduce the need to sample-check the finished product. Build the checking into the time estimate when you quote for the job, rather than treating it as something to do if the day allows.
What to keep on the file
Keep the schedule, the categories and reasons, the record of who reviewed and who approved, and a note that a tool was used to identify candidates. If a recipient asks why something was withheld, or a complaint follows, the schedule is the answer. A general recollection that the set was "checked" is not.
Where Alesis fits
Alesis reads a matter's papers page by page, so when it flags a passage it names the page the passage came from, and any page it could not read is flagged rather than skipped. It prepares drafts and lists for a qualified person to review and sign off; it does not file, serve or email anything, and applying redactions to a file remains work your firm does in its own document tools. Where it cannot find support for a point, it says what is missing instead of guessing. It assists qualified professionals and does not replace them.
