Summarising witness statements with AI and spotting inconsistencies
By the ChamberOne team 4 minute read
Three weeks before the hearing, you have eleven witness statements, two expert reports and a stack of exhibits. You know there are conflicts between the accounts, but finding them means reading everything twice and keeping a map in your head. This is a job where AI can help, because lining up passages by topic is tedious and mechanical. It is also a job where it can mislead, because the difference between 'I think it was around nine' and 'it was nine' is exactly what a cross-examiner needs. Here is a workflow that uses the tool for the first and guards against the second.
Key points
- Use AI to build an issue-by-issue index across statements, with paragraph references and quotations.
- Expect it to flatten hedges and to find both false links and false contradictions.
- Check every entry against the original before it goes into your working notes.
- Base cross-examination plans on the original words, and check the rules on AI and witness evidence.
What AI is good at here
A language model can read several statements and produce a table: for each event or issue, what each witness says. It can list the timeline each witness gives and show where timelines diverge. It can identify which witnesses say nothing on a topic, which is often as telling as what they say.
That gives you a starting index to the evidence. A paralegal might take a day to build it by hand. A tool can produce a draft in minutes, which you then check.
What it gets wrong
Summaries compress, and compression loses nuance. Hedges disappear. A witness who says 'as far as I recall' is turned into one who states a fact. A passage that is ambiguous gets resolved one way. Tone, hesitation and the order in which a witness chose to say things drop out entirely, though they may matter for credibility.
The tool may also invent a link. If two statements discuss the same meeting, it may treat them as describing the same moment when they do not. And it may smooth over a real contradiction because the two passages use different words for the same thing, or find a false contradiction because they use the same word for two things.
- Tabulates what each witness says on each issue
- Lists each witness's timeline and shows where they diverge
- Shows which witnesses say nothing on a topic
- Produces a draft index in minutes
- Drops hedges such as 'as far as I recall'
- Resolves ambiguous passages one way
- Invents links between statements
- Smooths over real contradictions or finds false ones
A workflow that holds up
- 1
Prepare the statements as clean text, with paragraph numbers kept, so every point can be traced back.
- 2
Ask for an issue-by-issue table, and tell the tool to give the paragraph number for every entry and to quote the key words, not paraphrase them.
- 3
Ask it separately for apparent inconsistencies between witnesses and within each witness's own statements.
- 4
Check every entry against the original paragraph. Mark each one confirmed, corrected or removed.
- 5
Only the confirmed table goes into your working notes.
- 6
Build cross-examination plans from the original words, with the table as an index.
Treat flagged inconsistencies as questions
An inconsistency flagged by a tool is a lead. Before you rely on it, ask whether it is real. Are the witnesses describing the same event? Did one speak from personal knowledge and the other from hearsay? Is there an innocent explanation, such as the passage of time or a different vantage point?
Also look for what the tool will not flag: consistencies that are suspiciously perfect. Statements that share the same phrases may suggest coordination, or may simply come from a drafting template. The tool will not tell you which.
Confidentiality and court rules
Witness statements are sensitive. Before you paste them into any tool, know where the text goes and who can read it. If the matter is restricted or involves protected identities, think carefully about whether to use AI at all.
Also check the rules on how statements are prepared. Some courts and practice notes restrict the use of AI in producing witness evidence itself. Using AI to analyse statements already made is a different thing from using it to write them, but you should read the applicable guidance and not assume.
A short scenario
Suppose a three-lawyer firm acts in a commercial dispute with eight witness statements. The associate has the tool produce a timeline table. It shows two witnesses placing a phone call before a delivery and one placing it after. She goes back to the originals and finds that one witness gave a time and the others only said 'earlier that day'. The apparent conflict shrinks to a matter for a gentle question rather than an attack. Without the check, she might have built a cross-examination around a contradiction that does not exist.
Preparing statements so the tool can read them well
Scanned statements often produce garbled text. Check the extraction before you ask for any analysis. A misread date or a missing 'not' changes the meaning and no later step will notice. Keep the paragraph numbers, headings and exhibit references in the text you supply, and keep each statement as a separate document so the tool does not mix up who said what.
Where a statement was translated, note that, and keep the original-language text. Subtle differences between two translations of the same words can look like inconsistency. Similarly, where a witness gave oral evidence earlier or a statement was amended, keep the versions apart and label them. The tool can compare versions of a single witness's account, which is a useful check, but only if it knows which is which.
Ask the tool to keep a list of what it could not read or could not place. A gap in its output is easier to deal with than a confident entry built on a damaged page.
What stays with the lawyer
Deciding which inconsistencies matter, how to put a point to a witness and when to leave one alone is advocacy. It depends on the judge, the witness, the case theory and the room. A tool can reduce the reading. It cannot decide what to do with it.
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
AI in Witness Bundle and Proof of Evidence Preparation: What to Automate and What Not To
A practical guide for litigators on which parts of witness bundle and proof of evidence preparation AI can handle, and which parts it must not touch.
6 minute readHallucinated citations: what court sanctions teach about checking AI output
The patterns behind fake authorities reaching courts, and the supervision and checking controls that would have caught them.
5 minute readBuilding a litigation chronology with AI
Extract dated events, link each to its source page and check for conflicts before the chronology is used.
5 minute read