Improper Reliance on A.I. + Failure to Use Litigation Support Experts + Lack of Transparency = Failure to Comply With “Meet & Confer” Order

The court didn't fault the AI. It faulted a process nobody could inspect.

In Lighthouse Healthcare Center, LLC v. National Fire & Marine Insurance (C.D. Cal., 2026 WL 314898), the court found that the parties had not complied with its July order. That order told them to meet and confer on ESI protocols for the OASIS and OnBase databases, with litigation support specialists taking part. Each side retained one, but their involvement ended after the first session. Defendants had flagged problems with AI-generated searches, yet the parties never worked through them together, and defendants wrote their own protocols without plaintiff's input. Defendants also said that 8,140 OnBase "Ops No Coverage" records could not be reliably searched. The court ordered a "do over" with a neutral expert. We worked from Michael Berman's post on EDRM, published in October 2026, which quotes the order at length. The text we had cut off partway through, so we haven't seen the full remedy.

What did the court say went wrong?

The parties treated the order as a one-off meeting. Specialists attended the start but were absent when protocols were built, AI search problems went undiscussed, and one side wrote its protocol alone.

The July order gave the specialists specific work. They were to assess what could be extracted from defendants' systems, explore what "monitor and adjust" functions cover, and develop protocols for downloading data into usable formats. The court found their roles were limited. Defendants' in-house specialist took instructions from counsel, and his declaration did not address how dataset problems were resolved or what methodology was used. Berman reads the ruling as saying plaintiff could not rely solely on AI, and that defendants should have been more transparent.

Is this a ruling against using AI in review?

No. As reported, the court objected to unsupervised reliance on a tool and to a lack of shared method, not to the technology. The missing piece was an expert-led process both sides could see.

One detail deserves attention. Defendants did tell plaintiff about the problems with the AI searches, and that still wasn't enough. Disclosing a limitation is not the same as resolving it under the order's terms. We have covered this pattern before in the WG13 "what AI's failures teach lawyers" piece. The sanctions-prone failure is usually the missing human process around the tool.

How much weight should buyers give the 8,140 figure?

Treat it as a party's own statement, not a measured error rate. The excerpt doesn't say who counted the records, how "reliably" was defined, or what share of the collection they represent.

Without the size of the underlying collection, 8,140 could be a rounding error or a serious gap. The same caution applies to any "coverage" or "accuracy" figure a vendor quotes. Ask who defined the metric, who measured it, and whether anyone independent can reproduce it. The court also noted that some files were too large for the AI dataset or were reported as errors. That is a reminder that processing exceptions are a core part of the record, not a footnote.

What should I ask my vendor now?

Ask for a complete exceptions report listing every file the AI could not process, the reason, and who dealt with it. Also ask for a written method an independent specialist can audit.

Specific questions follow. Which file types or sizes fail, and does the tool say so or silently skip them? Can your specialist stay involved through protocol design, not just kickoff? Can the platform export its search and extraction methodology for the other side or a neutral? A system that fails closed and itemises what it skipped is far easier to defend than one that quietly returns a clean-looking result. Given the insurer-pricing concerns we covered recently, expect carriers to ask these questions too.

Frequently asked questions

Who has to retain the specialist?

The July order had each party engage a litigation support specialist. The interim order also required an independent third-party ESI specialist, and the do-over uses a neutral expert.

Is this binding precedent?

It is a single district-court discovery-management order, reported through a blog post. Treat it as a warning about process, not a rule about AI.

Which AI tool was involved?

The excerpt refers to "Plaintiff's AI tool" and an "aiR assist dataset". We haven't seen enough of the order to say more, and we won't guess.

Sources: EDRM, "Improper Reliance on A.I. + Failure to Use Litigation Support Experts + Lack of Transparency = Failure to Comply With 'Meet & Confer' Order" (Michael D. Berman)