Auditing machine-generated deposition summaries
A testimony-preserving audit method that separates answers, assumptions, objections, uncertainty, and unsupported summary claims.
For: Deposition researchers and lawyers reviewing litigation-support output
Provide a claim-level method for checking whether a generated summary fairly reflects a transcript.
How should a machine-generated deposition summary preserve testimony and uncertainty?
Anchor each material summary claim to the transcript, retain the question that gives the answer meaning, and preserve qualifications such as uncertainty or lack of recollection. Separate what the witness said from an analyst's inference. Review contradictions and omissions as carefully as invented facts.
Identify the transcript before summarizing it
Federal Civil Rule 30 distinguishes witness review/change procedures and officer certification. A summary should therefore say which transcript version it used and whether related change material was considered. [S1]
- Witness identifier, deposition date, volume, reporter version, and page/line convention.
- Status: rough, certified, corrected, or otherwise described by the actual source.
- Exhibits available and unavailable; related errata or change statements.
- Summary author/tool, review date, and limits of the inspected record.
Tag claims by what the testimony supports
Use a proposed five-part vocabulary: direct testimony, qualified testimony, question premise, analyst inference, and unresolved conflict. This makes a summary auditable without implying that a witness adopted every premise embedded in a question. A 'yes' may need the full question and preceding clarification to be intelligible.
- Direct testimony: a supported statement with an exact locator.
- Qualified testimony: preserve 'approximately,' 'I think,' or similar limits.
- Question premise: do not treat an attorney's assertion as the witness's fact.
- Inference: explain the reasoning and label it as analysis.
- Conflict: link both passages without silently selecting one.
Synthetic transcript audit: memory is not a denial
Fictional transcript, page 12 lines 3–7: Q: 'Did you approve shipment on May 4?' A: 'I do not remember approving it. I usually review shipping requests on Fridays.' A defective summary says, 'Witness denied approving the May 4 shipment.' The supported summary says, 'Witness did not recall approving the shipment and described a usual Friday review practice.'
- The date in the question is a premise, not independently confirmed testimony.
- The answer reports a memory limit, not a categorical denial.
- The usual practice does not establish what happened on this occasion.
- Any inference about timing remains separate and requires additional support.
Audit both additions and omissions
A useful check reads surrounding testimony, samples uncited passages for omitted qualifications, and examines whether the summary overstates certainty. ABA Formal Opinion 512 addresses independent review appropriate to the tool and task. The procedure here is a proposed way to operationalize review, not an ABA-mandated checklist. [S2]
- Check every quotation and every outcome-significant characterization.
- Look for missing objections, corrections, incomplete answers, and exhibit dependencies.
- Record reviewed claims, rejected claims, and unresolved passages separately.
- Release the summary with its source scope and review status; retain access to the underlying transcript.
Scope and limitations
Educational audit method with a wholly synthetic transcript excerpt. Federal civil deposition rules are a reference point, not a substitute for governing rules, orders, or agreements.
No testimony, witness credibility, admissibility, or actual deposition record was evaluated. The example is invented.
Sources
- Federal Rules of Civil Procedure — December 1, 2025 compilation — Federal civil rules, particularly Rule 30 on depositions and review/certification. Checked 2026-11-05
- American Bar Association — Formal Opinion 512 (2024) — ABA interpretation of Model Rules for generative AI, including competence, confidentiality, supervision, communication, and fees; local law must be checked. Checked 2026-11-05
Explore further
This audit method connects to human review of AI-assisted legal work. Visit Agentic Attorney at https://agenticattorney.com/ for the broader workflow context; this page does not host transcripts or offer court reporting. Explore Agentic Attorney.