Generative AI Document Review After Schulte v. LinkedIn: The Defensible Workflow

On 3 June 2026 the Ninth Circuit suspended two lawyers from practising before it for six months over AI-fabricated citations, writing that “a competent and diligent attorney must also read and reason”. Four weeks later, on 1 July 2026, a federal court in the same circuit looked at a defendant that had let generative AI make responsiveness calls across its document population and denied every motion the plaintiffs brought against it.

Same technology, same month, opposite outcomes, and the difference is not the model. Generative AI document review won in Schulte v. LinkedIn because it ran over a closed set of documents, behind search-term culling, with human sampling in every category and a record of how it was done. Generative research keeps losing because “add case law” runs over an open world with none of that.

The ruling: aiR treated as “a form of technology-assisted review”

Schulte v. LinkedIn Corp., No. 22-cv-00237-HSG (LB) (N.D. Cal., 1 July 2026), is a case in which LinkedIn produced documents in two stages: 25 search strings, then Relativity aiR for Review making responsiveness calls with no training set, validated by human sampling per category. The plaintiffs brought three motions; all three were denied.

The order itself is not on RECAP at the time of writing; what is known comes from DLA Piper’s summary, and it carries three holdings. Generative AI review is “a form of technology-assisted review”, so existing TAR case law applies rather than a new standard. Search-term culling before the AI pass was reasonable and proportional. And demands for LinkedIn’s validation metrics were “discovery on discovery”, requiring a specific showing of deficiency.

Pre-culling with search terms is reasonable and proportional

One objection was to the two-stage design: keywords first, AI second. The court’s answer, as summarised, was that “courts have held that using search terms to pre-cull documents before technology-review platforms satisfies the reasonableness and proportionality standards”.

The largest cost and risk in an AI review is the size of the population you run the model over; Reed Smith describes aiR’s sweet spot as populations of 50,000 to 100,000 documents; culling is how a million-document collection becomes one, and you lose no defensibility by doing so.

“Discovery on discovery”: why the plaintiffs did not get the validation metrics

The plaintiffs also sought LinkedIn’s validation metrics: the sampling results behind the responsiveness calls. Denied, because a requesting party must show a specific deficiency before it gets to interrogate the other side’s process.

That cuts both ways. You are not obliged to hand over your metrics on request; you are obliged to have them, because the moment the other side finds a responsive document you missed, the deficiency showing is made and the question becomes “what did you do to validate?” Schulte protects the record; it does not excuse you from making one.

Reed Smith’s validation protocol: 50-100 seed documents, 500-1,000, full population

Reed Smith has published the most concrete account of how a large firm validates aiR, and it is the protocol a court will recognise: test the review prompt on 50 to 100 seed documents a lawyer has already coded; analyse every disagreement; expand to 500 to 1,000; run the full population with sampling throughout; document every prompt. The firm reports internal investigations with under 1% relevant material done “in hours instead of months”. The prompt below drafts that protocol for your matter; the shape matters, the numbers are yours to defend.

Draft a validation plan for our AI-assisted responsiveness review of [N] documents in [matter], run in [Relativity aiR / Everlaw AI Assistant]. Steps: (1) seed set of [50-100] documents already coded by a lawyer, reviewed by the tool, with a disagreement analysis by category; (2) expansion to [500-1,000] documents with random sampling at [rate]; (3) full-population run with ongoing sampling at [rate] and a hold-out set; (4) a prompt log recording every prompt version, date and author; (5) an escalation rule for any category with an error rate above [X%]; (6) the record we will keep and who signs it. Output a one-page protocol plus a table of what gets logged, by whom, and when.

The review criteria are the second prompt, and where the skill lives: write them as instructions to a first-year reviewer, not as a legal test.

You are coding documents for responsiveness in [matter]. A document is RESPONSIVE if it concerns any of the following, each tied to a request: [Request 3: communications between [Custodian A] and any [Company B] employee about [the pricing change] between [date] and [date]]; [Request 7: ...]. It is NOT RESPONSIVE if it concerns none of them, including [routine newsletters, calendar invitations without substantive content, personal messages]. Mark NEEDS REVIEW where a listed topic is mentioned only in passing, the document is not in English, or it is a fragment. For every RESPONSIVE or NEEDS REVIEW call, quote the words that triggered it and name the request. Do not assess privilege; that is a separate pass.

Write the criteria with the lawyers who coded the seed set, rerun that set after every prompt change, and log the disagreements, not just the agreement rate; the disagreements are what you will be asked about.

What Relativity and Everlaw now include for free

The economics changed at Relativity Fest in October 2025, where CEO Phil Saunders called generative AI “the undeniable future of review” and Relativity announced that aiR for Review and aiR for Privilege would be included in standard RelativityOne pricing at no additional charge from early 2026 (aiR for Case Strategy stayed separate), reporting 200+ customers and 25 million documents reviewed; its headline case, a vendor figure, was Purpose Legal cutting review time by 85% on a 300,000-document matter. Everlaw likewise made single-use AI Assistant features free. ILS called the two moves landmark because “deep-pocketed defendants have been able to outspend plaintiff firms on document review”.

One caution on the tool race. The SKILLS Legal AI Survey of 130 AI leaders at the largest firms (Oz Benamram, March 2026), published on Harvey’s own blog, found incumbents still dominate e-discovery even as Harvey leads in “discovery automation” and “timelines and chronologies”. The review platform is where the documents sit, the audit trail exists and the protective order applies; use the bake-off guide to choose between platforms, not to leave one.

Prompt documentation as the audit trail

In a TAR review the audit trail was the training set and the statistics; in a generative review it is the prompt history plus the sampling. Reed Smith documents every prompt, and one governance tracker lists a usage log among the artefacts insurers ask for at renewal. The minimum record:

Ask the model at the end of each session to summarise it as a log with those fields, then initial each item verified or not; never mark verified what nobody has checked. More review prompts are in the prompt library.

Privilege review and the 80% time reduction claim

Relativity says aiR for Privilege delivers up to 80% time reduction. It is a vendor number, and privilege is where a vendor number deserves the most caution: a privilege log is a representation to the court, and a missed waiver flag is not a statistic. The tool proposes, the lawyer decides.

For each document in this batch, propose a privilege log entry: Bates | Date | Author | Recipients (mark lawyers with *) | Document type | Privilege claimed (Attorney-client / Work product / Both / None apparent) | Basis in one neutral sentence that reveals no privileged content | Confidence (High/Medium/Low). Flag separately: any document sent to a third party (possible waiver); any where no lawyer appears; any that reads as business rather than legal advice. Do not summarise the privileged content. Where you cannot determine a field, write UNKNOWN.

A lawyer reads every waiver flag, every “Both” and “None apparent”, and every basis line, because a description that discloses the advice defeats the log. Whether the tool itself can compromise privilege turns on its contractual terms, as the privilege guide explains.

Where the documents may go: the protective orders

Two 2026 orders decide which tools are usable for discovery material at all; both are summarised by Akin Gump. In Jeffries v. Harcros Chemicals (D. Kan., 25 March 2026) the court banned public “open” AI tools for all discovery materials, confidential or not, because processed data is “practically impossible” to claw back. In Morgan v. V2X (D. Colo., 30 March 2026) AI platforms are permitted only where the provider is contractually prohibited from “(1) storing or using inputs to train or improve its model; and (2) disclosing inputs to third parties except where essential”, which the court acknowledged “practically bars the use of most ‘low-to-no-cost’ AI tools”.

Why generative AI document review wins where generative research loses

The contrast in the opening tells you which other AI tasks will survive judicial scrutiny. Review is a closed-universe task: the documents are in the tool, the criteria are written down, the output is checkable by sampling; Vals’ 2025 benchmark scored document Q&A, the closest cousin, at 94.8% for the best tool against 70.1% for lawyers. Research is an open-world task: Stanford’s 2024 study found even the paid platforms hallucinating on more than 17% of queries, and the sanctions record from Mata to Lnu is a record of lawyers filing what they had not read. The guide to AI legal research without hallucinations covers that side of the line.

The discipline that makes review defensible (a closed set, written criteria, sampling, a log) makes any AI output defensible. That is the mindset AI Lab for Lawyers teaches, on discovery material and everything else: sample, document, sample again.

Checklist for a defensible AI review

1. Negotiate the protocol first: culling terms, that generative review will be used, what will be disclosed.
2. Check the protective order for Jeffries- or Morgan-style AI clauses and confirm the platform’s terms satisfy them.
3. Cull with agreed search terms; record the population before and after.
4. Write criteria with the lawyers who coded the seed set, one request per criterion, with a NEEDS REVIEW exit.
5. Validate in three stages (50-100, 500-1,000, full population), sampling throughout, rerunning the seed set after every prompt change.
6. Log disagreements, escalations and sign-off; run privilege as a separate pass with a lawyer reading every flag.
7. Keep the record as long as anyone might ask for it.

Where to go next: the case chronology guide and the deposition summary guide cover what happens to the documents once they are in; the use-case hub, the guide to how lawyers use AI day to day and the litigation practice-area guide put review alongside the rest of the toolkit; and the best AI tools for lawyers overview sets the e-discovery products in context.