AI review
What makes an AI discovery summary verifiable?
A polished summary is only useful when the reviewer can check the source, the context, and the limits of the work.
On this page
A verifiable AI discovery summary gives a reviewer a specific source, a precise location, and an honest account of uncertainty. It also makes clear which files were covered. The key test is whether another person can check a finding without reproducing the entire review.
A source reference should be usable
A filename alone can still leave a reviewer searching through hundreds of pages. Look for page references or recording timestamps and a practical way to open the source. Check that the referenced file is the version actually reviewed.
When a finding draws on several sources, each source should have a clear role. One might establish a date, another a person’s account, and another an unresolved inconsistency. A list of citations at the end of a paragraph does not automatically establish every sentence in it.
Exact quotes and summaries need different checks
An exact quotation should match the source. A summary should preserve the meaning and identify whose account it describes. Neither should hide missing context. Reopen the passage around a consequential quotation and check whether a qualification changes its meaning.
For recordings, compare the transcript with the audio. For scans, compare extracted text with the page image when a key word or number matters. A correct-looking citation can still point to incorrectly extracted text.
Coverage belongs beside the result
Ask what happened to every submitted file. A result based on ten readable documents is not a complete account of a folder containing those documents and several unprocessed recordings. A review tool should help the firm recognize that difference.
Missing support also needs a scope. “No support located in the reviewed files” is a review observation. It does not establish that evidence does not exist elsewhere, or decide the legal sufficiency of a charge.
When did the stop begin?
- Finding
- The report places the stop at 9:14 p.m.; the dispatch log records the call at 9:16 p.m.
- Sources
- Arrest report, page 1 (exact quotation) · Dispatch log, page 3 (summary)
- Coverage
- Based on 10 of 13 files. Three recordings could not be processed and were not reviewed.
Arrest report, page 1“At approximately 2114 hours I initiated a traffic stop.”
Compare workflows, not just answer quality
A general chat tool may help explain a passage or draft working notes from an authorized sample. For ongoing discovery review, also evaluate source navigation, file coverage, case access, corrections, and the handling of new versions. Capabilities vary by product and configuration, so test the actual tool your team would use.
A small evaluation can include a known quotation, a misleadingly similar passage, a missing file, and an updated version. Check whether the tool distinguishes these cases. Record failures as well as successes; one strong answer is not a complete evaluation.
Keep attorney review explicit
Florida Bar Ethics Opinion 24-1 discusses lawyers’ responsibilities when using generative AI, including confidentiality and checking work for accuracy. It is an advisory opinion. Firms should assess its application to their own practice.
DiscoRev is built to assist review, not make final legal judgments. Its beta focuses first on Florida criminal-defense firms. The discovery review checklist shows how to keep AI-assisted work connected to a broader human review process.
The printable checklistFive review steps on one page. No email required.OpenCommon questions
How do you check an AI-generated discovery summary?
Open the cited source at the referenced page or timestamp, confirm that exact quotations match, read the surrounding passage for context, and check which files the summary actually covered.
Can AI decide whether a charge is supported?
No. An AI tool can point to support and gaps in the files it reviewed. The responsible lawyer decides legal sufficiency and what the evidence means.