Yes. In Schulte v. LinkedIn, 2026 WL 1905851 (N.D. Cal. June 30, 2026), Magistrate Judge Laurel Beeler treated LinkedIn’s generative AI review (Relativity aiR) as technology-assisted review, not as a new species of discovery that needs its own rules.
The court applied the parties’ existing ESI protocol and ordinary Rule 26 reasonableness and proportionality. It did not bless a product, and it did not create a national “GenAI is TAR” statute.
What the order actually held
Plaintiffs challenged LinkedIn’s use of a generative AI tool to make responsiveness calls after keyword culling. The court measured that workflow the way courts have measured TAR since Da Silva Moore and Rio Tinto: is the process reasonable and proportional, and did the producing party do what the ESI protocol required?
It declined to compel extra “discovery on discovery” — elusion rates, error rates, reviewer counts — without a concrete showing that the production was deficient. Disclosing that there was no seed set, that the tool made final responsiveness calls, and that humans sampled for QC was enough on that record.
What it does not do
It is one magistrate order in the Northern District of California. It is not a holding that every GenAI vendor is validated TAR, or that you can skip validation, human review, or Rule 26(g) certification. Permission to use a method is not a finding that your use of it was reasonable.
If your ESI protocol says “TAR,” assume a court may read GenAI into that word. Write the protocol for that outcome: what you will disclose, what you will not, and how you will validate.
How to use the case tomorrow
If opposing counsel says GenAI is untested, Schulte is the cite that it is another predictive engine in a familiar TAR workflow. If they demand elusion estimates and model cards as a matter of right, the order says that is discovery-on-discovery unless they first show a hole in the production.
If you are the producing party, do not read it as “say nothing.” LinkedIn disclosed the tool, the no-seed-set design, the final-call posture, and the sampling plan. Keep that file. Pair it with the longer defensibility checklist: tool, prompts, validation, humans, exclusions, and what you told the other side.