Can AI Summarise a Deposition Accurately? How to Get Page-Line Cites You Can Check

Yes, AI can summarise a deposition with page and line cites, but only if you give it a transcript where every line carries its page and line number, and then check each cite with a script and by eye. Without numbered text, models produce plausible-looking cites that point at the wrong passage.

The summary is the easy part. The cites are the whole value of a digest: a lawyer using it will jump straight to "45:11-13" and expect to find what the digest says. The page-line test is whether those cites survive a check, and you can run it on your own transcript: prepare the text, prompt for cites, verify them with a script, and settle under the privilege rules where the work may happen.

Follow me on Instagram@sagnikteaches

What a page-line digest must contain

A digest is a chronological or topical list of what the witness said, each entry tied to the exact place in the transcript. A usable entry has four parts: the cite, who said it (witness or questioner), a neutral summary of what was said, and a flag if the answer is a hedge, a contradiction or an objection. The neutral wording matters. "Witness admits the worktop was fitted by someone else" is an argument; "Witness says brother fitted the worktop at the weekend" is a summary.

Connect on LinkedInSagnik Bhattacharya

Digests come in two shapes. A page-by-page digest follows the transcript order and suits a quick refresher. A topic digest groups testimony by issue (for example, who did the fitting, payment, notice of defects) and suits trial preparation. You can ask the model for either from the same prepared text.

Subscribe on YouTube@sagnikteaches

Why PDFs defeat the cite

A certified transcript usually arrives as a PDF with a fixed number of numbered lines on each page. When the PDF is converted to text, the line numbers may land at the start of each line, appear in a block at the page edge, or vanish; page numbers may sit in a header the extractor drops. The model sees a stream of words with no reliable coordinates and fills the gap from habit. The outcome is cites that look right (page 45, lines 10 to 14) while pointing a few lines or a whole page away from the real passage.

So prepare the text before any summarising. The aim is one line of text per transcript line, each starting with a bracketed marker. Here is a small script that converts a plain-text export where pages begin with "Page N" and lines start with their number. Adapt the two patterns to your transcript's layout, and change the 25 in the line-number test if your transcripts carry a different number of lines per page.

import re, sys
page = None
out = []
for raw in open(sys.argv[1], encoding="utf-8"):
    line = raw.rstrip("\n")
    m = re.match(r"^\s*Page\s+(\d+)\s*$", line)
    if m:
        page = int(m.group(1))
        continue
    m = re.match(r"^\s*(\d{1,2})\s+(.*)$", line)
    if page and m and 1 <= int(m.group(1)) <= 25:
        out.append(f"[{page}:{int(m.group(1))}] {m.group(2)}")
open(sys.argv[2], "w", encoding="utf-8").write("\n".join(out))

On a short test excerpt the output looks like this, and it is the form you paste to the model:

[45:10] Q.  Who fitted the worktop?
[45:11] A.  My brother did, at the weekend.
[45:12] Q.  Did you pay him?
[45:13] A.  I gave him cash, about four hundred.
[46:1] Q.  Was there an invoice?
[46:2] A.  No, there wasn't one.

If your transcript arrives as a scan, run OCR first and then read a few pages against the paper, because OCR errors in a line number break a cite as surely as a model error does. The court reporter tutorial explains how transcripts are produced, which helps when you need to ask the reporter for a clean text file (ask whether one comes with the PDF).

The digest prompt

You are preparing a page-line digest of a deposition for a litigation team.
The transcript below has a [page:line] marker on every line.
Rules:
1. Write one entry per topic shift or significant answer. Start each entry
   with the cite in the form [45:11-13] or [45:25-46:3], using only the markers
   in the text.
2. Say who spoke. Never attribute a question's content to the witness.
   Summarise only what the witness actually answered.
3. Put exact words in quotation marks when a phrase matters, copied
   character for character from the text.
4. Flag each entry as: PLAIN, HEDGED (witness unsure), OBJECTION, or
   POSSIBLE CONFLICT with an earlier cite (name the cite).
5. If you are unsure of a cite, write CITE UNSURE instead of guessing.
6. Do not characterise credibility. Do not infer motive.
TRANSCRIPT:
[paste prepared text, one chunk of up to about 60 pages]

Chunking matters. A 300-page transcript runs to tens of thousands of words. Some plans will accept that much in one go, but a single long pass makes a wrong cite harder to spot and harder to trace back to the passage that caused it. Feed it in chunks of about 50 to 60 pages, with the chunk boundaries on a natural break, and ask for a digest per chunk. Then ask for a separate short pass that compares chunks for conflicts, using the digests rather than the raw text.

An illustrative entry from a chunk for a kitchen fitter dispute, where a homeowner is the deponent:

[45:10-13] Witness says his brother fitted the worktop, over a weekend, and that he paid "about four hundred" in cash. PLAIN
[46:1-2] Asked about an invoice, witness says there was none. PLAIN
[47:5-9] Witness unsure when he first noticed the crack: "maybe March, maybe April". HEDGED

The cite check: let a script do the dull part

Once you have a digest in that bracket format, a second script reads each cite, fetches the cited lines from the prepared text and tests whether every quoted phrase in the entry really appears there. It does not judge meaning; it catches invented or misplaced quotes and cites that point at no line.

import re, sys
from difflib import SequenceMatcher

lines = {}
for l in open(sys.argv[1], encoding="utf-8"):
    m = re.match(r"\[(\d+):(\d+)\] (.*)", l.strip())
    if m:
        lines[(int(m[1]), int(m[2]))] = m[3]

def span(p1, l1, p2, l2):
    keys = sorted(k for k in lines if (p1, l1) <= k <= (p2, l2))
    return " ".join(lines[k] for k in keys).lower()

for d in open(sys.argv[2], encoding="utf-8"):
    m = re.match(r"\[(\d+):(\d+)(?:-(?:(\d+):)?(\d+))?\]\s*(.*)", d.strip())
    if not m:
        continue
    p1, l1 = int(m[1]), int(m[2])
    p2 = int(m[3]) if m[3] else p1
    l2 = int(m[4]) if m[4] else l1
    text = span(p1, l1, p2, l2)
    for q in re.findall(r'"([^"]+)"', m[5]):
        q = q.lower().strip(" .,")
        if q not in text:
            r = SequenceMatcher(None, q, text).find_longest_match(0, len(q), 0, len(text)).size / max(len(q), 1)
            print(f"CHECK {m[1]}:{m[2]}  quote not found verbatim (overlap {r:.0%}): {q[:60]}")
    if not text:
        print(f"CHECK {m[1]}:{m[2]}  cite points at no line")

Run as python verify.py prepared.txt digest.txt. On the test excerpt above (plus a few lines from page 47), a digest entry for 46:1-2 that put "an invoice from the firm" in quotation marks printed CHECK 46:1 quote not found verbatim (overlap 42%): an invoice from the firm, and an entry citing 47:3, a line that does not exist in the file, printed CHECK 47:3 cite points at no line. Those are exactly the faults you want surfaced before a lawyer relies on the entry. The script cannot confirm that your summary is a fair paraphrase: a person still reads the lines behind any cite that matters.

A 212-page run for a kitchen fitter dispute, illustrated

An illustration with assumed numbers. A small litigation team has a 212-page deposition of a homeowner in a dispute with a kitchen fitter. They prepare the text, split it into four chunks of 53 pages, and run the prompt, producing 62 entries. The script flags 9 entries. Reading the lines behind those 9, the team finds: 4 quotes with a changed word (a paraphrase placed inside quotation marks), 3 cites one line off, 1 cite one page off, and 1 quote that does not exist anywhere in the transcript. After correction, a reviewer reads the passages behind a random tenth of the other entries (6 of 53). All six are accurate.

The sums: 9 flagged of 62 is about 15%, which is the reason nobody should paste AI cites into a filing unchecked. Roughly 1 in 62 entries (the quote that did not exist) was a fabrication that would have embarrassed the team. The team's time: about 20 minutes to prepare the text, 40 minutes to run the chunks, 30 minutes to check flagged entries and 30 to read the sampled passages, against what they estimate at 6 hours for a paralegal's first manual digest. These figures are assumptions for the illustration; time your first transcript properly before you quote a saving to anyone.

Four failures a reviewer should expect to find

The question taken for the answer. Counsel asks, "So you told him the crack was already there before the fitting, correct?" and the witness answers "No." A digest saying "witness told him the crack pre-dated fitting" reverses the testimony. It showed up when a reviewer read the lines behind a cite flagged POSSIBLE CONFLICT. The prompt rule "never attribute a question's content to the witness" is there for this.

The cross-page entry. An answer starts on 61:24 and ends on 62:3. The model cites only 61:24-25, so the digest omits the sentence that changes the meaning. Ask for cross-page cites in the form [61:24-62:3] and spot-check any entry that ends on line 24 or 25.

The errata sheet. The witness later corrects an answer on a signed errata sheet. A digest built only from the transcript does not know. Keep a separate list of corrections, and add each to the digest as its own entry citing the errata page.

The harmless-looking paraphrase. "Witness was vague about payment" is an inference. The check script cannot catch it. Your reviewer's rule is that any adjective describing the witness's state of mind or credibility comes out of the digest.

Other uses once the text is numbered

• Contradiction search. Ask for all statements about a specific fact, such as "who fitted the worktop", with cites, across the whole transcript. Then read the lines.
• Exhibit index. List every exhibit mentioned, the cite where it is first shown, and what the witness said about it.
• Objection log. List each objection with the cite, the stated ground and whether the witness answered. Check the grounds against the lines; models misread shorthand.
• Follow-up outline. Draft questions for a later deposition or trial cross, each linked to the cite that prompts it. A lawyer owns the strategy; the model only organises.

A specialist food importer's contract dispute, a childcare nursery's employment claim and a craft brewery supplier's supply-agreement case all use the same preparation. What changes is the glossary: give the model a short list of names, product terms and defined terms, so "Mr [surname]" and "the buyer" are linked to the right party.

A topic digest, with a sample output

For trial preparation, a topic digest is often more useful than the page-by-page version. Run it after the first digest is checked, and feed it the checked entries rather than the raw transcript, so every cite it uses has already passed the script.

Below is a checked page-line digest. Group the entries under these issues:
(1) who fitted the worktop, (2) payment, (3) when the defect was noticed,
(4) contact with the fitter after the defect.
For each issue list the relevant cites in order with a one-line neutral summary.
Use only cites that appear in the digest. After each issue, list any pairs of
cites that appear to conflict, naming both cites, and say nothing else about them.

An illustrative result for issue (3): "[47:5-9] Unsure when he first noticed the crack, 'maybe March, maybe April'. [52:14-18] Says he saw the crack 'the week we moved the fridge'. [88:3-6] Says he emailed the fitter on 2 April. Possible conflict: 47:5-9 and 88:3-6, on timing." What you fix: the quoted phrases must match the transcript (the script checks that), and "possible conflict" is a pointer for a lawyer to read, not a finding. Only a lawyer decides whether "maybe March" and an email on 2 April are in tension.

Filling in the glossary that stops confusion

Depositions are full of shorthand, nicknames and defined terms. A short glossary at the top of each chunk prevents the model treating one person as two or two as one. A filled-in example for the kitchen fitter dispute:

GLOSSARY (use these names in summaries):
- Witness = the homeowner (deponent)
- "my brother" = the witness's brother, who fitted the worktop
- "the firm" / "Fitter Co" = the defendant kitchen fitter
- "the island" = the central unit in the kitchen
- Exhibit 4 = the quotation dated in January; Exhibit 7 = the March email
Questioner A = counsel for the defendant; Questioner B = counsel for the claimant

Speaker labels matter in multi-lawyer depositions. If the transcript just says "Q." for every question, the model cannot tell which side asked, and summaries of friendly cross-examination look like hostile questioning. Add the questioner's name to the prepared text at each change of questioner, using the transcript's appearance page.

An objection log, checked

Objections are short and shorthand-heavy ("Objection, form"; "Asked and answered"), which suits a table. Ask the model for one row per objection and check the cites with the script, because the words quoted in each row are short enough for the verbatim test to work.

The "answered?" column is where a model is most likely to err, since it needs the next few lines. Read each of those on the page. Any "No" row is worth a lawyer's attention at once because it may need a motion.

Video depositions and speaker tags

If the deposition was recorded, the transcript is the record, but a video sync file may exist with a timestamp for each page and line. Where it does, keep the timestamp in the prepared text too (for example [45:11 | 00:12:44]) so each digest entry can point both to the page and to the clip. Do not use a general note-taking tool's own automatic transcript of the recording as a substitute for the certified transcript: its words are not the record and its line numbers do not exist.

Keeping chunks consistent

When you digest in four chunks, the entries in chunk three do not know what the witness said in chunk one. Three habits keep the whole consistent. First, give every chunk the same glossary and the same prompt, unchanged. Second, ask each chunk's digest to keep issue labels from a fixed list you supply, so that "payment" is never "money" in one chunk and "fee" in another. Third, run the conflict pass only on the digests and cites, then read the lines behind each flagged pair. If a chunk boundary falls in the middle of an answer, overlap the chunks by about two pages and delete the duplicate entries afterwards.

One answer, a weak digest entry and a checked one

Here is a short passage, then two ways it can be summarised. The transcript reads: [71:3] Q. You told the fitter about the crack in April? [71:4] A. I think I emailed him, but I can't say the date. [71:5] Q. Did you phone him? [71:6] A. No.

The weak entry reads "[71] Witness told the fitter about the crack in April by phone and email." It has no line numbers, takes the date from the question, and invents a phone call the witness denied. The checked entry reads "[71:3-6] Asked whether he told the fitter about the crack in April, witness says he thinks he emailed but cannot give the date, and says he did not phone. HEDGED". It names the lines, keeps the hedge and records the denial. The difference is entirely in the prompt rules and the review: question content is never testimony, hedges are flagged, and cites carry line numbers.

A short habit helps a whole team. Pick three entries at random from every digest before it leaves the matter team, open the transcript beside them, and write "checked" with initials in the digest header. It takes five minutes, and it builds the habit that makes the larger checks happen.

Cost and time in a small firm

For a two-person litigation team on a business plan, the spend is two seats at $20 a month billed annually ($40), or $50 monthly. The scripts cost nothing. The real cost is reviewer time: the figures in the test above put a 212-page transcript at about two hours of staff time instead of six, but only if the check steps stay in. Skip them and the saving is false, because the first wrong cite in a filing costs more than the whole workflow saves. Time your own first run and keep the log in the matter file.

Privilege and where the work happens

A deposition transcript is not usually privileged by itself, but your digest and notes often are (they reveal what counsel thinks matters), and the transcript may be subject to protective orders or confidentiality designations. Handle it in this order.

1. Check the order. If a protective order restricts who may see the transcript or where it may be stored, read what it says about third-party tools before you upload anything.
2. Use only a business or enterprise plan. ChatGPT Business and Claude Team list at $20 per user per month billed annually ($25 monthly), with two seats minimum, and business content is not used for training by default on either plan. A free or personal account is off limits for client matters.
3. Check consent and ethics guidance. Engagement letters and your bar's guidance on generative AI may require disclosure or consent. Ask your general counsel or ethics partner; this tutorial is not legal advice.
4. Mark the output. Label digests as privileged and confidential, prepared for counsel, with the transcript version (rough or certified) and the date.
5. Keep it out of shared chats. Do not use a shared team chat where members outside the matter can see it, and delete working chats under your retention policy once the digest is filed.

The same caution applies to the cite check scripts: they run on your own machine, so they add no new disclosure. The wider question of which legal tasks to keep away from AI is in the tasks a firm should never hand to AI, and the comparison of ChatGPT and legal AI tools helps if you are deciding between a general chatbot and a specialist product. For the general method of summarising long documents safely, see summarising bundles and transcripts with AI.

A short acceptance test before you adopt it

Run this on one finished matter whose digest a paralegal already wrote.

1. Prepare the transcript text and confirm 10 random lines against the PDF.
2. Generate the digest in chunks and run the cite check.
3. Compare the AI digest with the human digest on three issues: does it find the same key admissions, does it include anything the human missed, and does it contain any statement the transcript does not support?
4. Count the cites the script flagged, and the cites a human found wrong after the script passed.
5. Decide the rule: for example, "AI digests are used for first-pass review only; any cite quoted in work product is verified by a lawyer against the certified transcript."

If the second flagged count is more than zero or two in a typical transcript, tighten the rule. Treat the digest as a map, never as the testimony.

Questions litigation teams ask about AI digests

Why do AI deposition summaries get page and line cites wrong?

Most PDF exports bury line numbers inside the text or drop them, so the model has nothing reliable to cite and may reconstruct numbers from memory. The fix is to convert the transcript to text with a visible page and line marker on every line before the model sees it, then verify each cite against that text.

Is it safe to put a deposition transcript into ChatGPT or Claude?

Only on a business or enterprise plan where content is not used for training by default, with client consent where your engagement terms or ethics rules require it, and under your firm's written policy. Do not use a free or personal account. Check your bar's guidance on generative AI before starting.

How many cites should I check by hand?

Run the script on all of them, then read the passages behind every cite that will appear in a filing or a witness outline, plus a random tenth of the rest. If the random tenth finds more than one or two problems, check everything.

Can AI summarise a rough or uncertified transcript?

Yes, but page and line numbers can shift between a rough draft and the certified version. Mark every digest with the version it was built from and re-run the cite check when the final transcript arrives.

Further reads

• How Small Law Firms Use AI to Draft Letters and Routine Documents — Routine letters a small firm can draft with AI.
• AI Paralegal Software: What a Small Firm Should Test Before Paying — What to test before paying for paralegal-style AI software.
• AI for Small Law Firms: What to Automate First — Which jobs a small firm should automate first.
• Best Legal AI for Criminal Law: A Small Defence Firm's Shortlist — A defence firm's shortlist of legal AI.
• AI Implementation Plan for a Small Law Firm: The First 90 Days — A first-90-days plan for a small firm adopting AI.
• Law Firm AI Usage: How to Audit What Your Fee Earners Really Use — Find out what your fee earners already use.
• AI Legal Brief Writing: A Draft and Cite-Check Routine — A routine for small litigation firms: build an authority table, draft only from it, cite-check each source line by line, and meet court AI-disclosure rules.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: OpenAI ChatGPT Business and Claude pricing pages (prices, seat minimums and no training on business content by default, checked 4 November 2026); both scripts run locally on a synthetic test transcript. No other vendor claims are relied on.