📝 Show Your Work: A Judge Tells Meta “I Don’t Trust AI,” a Court Orders a Discovery Do-Over, and Experts’ AI Prompts Land in Discovery

Or: Meta’s AI Privilege Review Request Rejected, AI Searches Without Experts Trigger a Neutral-Expert Do-Over, Expert AI Prompts Ruled Discoverable and California’s SB 574 Reaches Review Teams: The Week’s eDiscovery AI News in Plain English

By Thomas Seest, Dunlap.Today · Thursday, October 22, 2026

Remember the math teacher who marked your answer wrong even though it was right, because you didn’t show your work? It felt unfair at the time. But she had a point: if she couldn’t see how you got there, she couldn’t tell whether you knew it or got lucky.

This week, judges sounded a lot like that teacher. One federal judge told Meta she wouldn’t let it use AI to redo a big privilege review unless all of the AI’s training and testing was disclosed. Another court ordered a do-over after a party leaned on AI searches without bringing in the tech experts the court had required. And a new survey of court decisions shows that expert witnesses’ AI prompts are increasingly something the other side gets to read.

Here’s this week’s roundup: why judges want to see AI’s work, what happens when AI searches skip the experts, expert witnesses and their AI prompts, a security warning for law firm backups, and a few tools and numbers worth knowing.

⏱️ What this means for you in 30 seconds

• Want to use AI on a contested review? Be ready to explain, and maybe disclose, how it was trained and tested.
• Court told you to involve tech experts? Keep them in the room for every meeting, not just the first one.
• Hiring an expert witness who uses AI? Assume their prompts can be requested, and plan for it in writing.
• Practicing in California? SB 574 takes effect Jan. 1. Check AI-drafted discovery responses and privilege logs before they go out.
• Using a Microsoft 365 backup service? Find out how long your sign-in and audit logs are kept.

🧢 “I Don’t Trust AI”: Meta’s Privilege Review Stays Human

In the big social media addiction case in federal court in California, Judge Yvonne Gonzalez Rogers wasn’t impressed by how many documents Meta had marked as privileged. According to the New York Post, she pointed out that Meta employees had even made caps about attorney-client privilege: “I have never seen a company whose employees are so taken with the attorney-client privilege that they’d have caps made.”

When a Meta lawyer asked to use AI for the re-review, the judge said: “I don’t trust AI, obviously, as much as you do… The only way that argument would have any legs is if you trained a model on your own, and obviously all of the training would have to be done and disclosed — all of it — and all of the testing.”

Under an agreement the court approved this week, Meta will re-review about 170,000 privilege claims at roughly 20,000 documents a week through Dec. 30. A special master, retired Judge Suzanne Segal, will spot-check 200 documents a week, and Meta pays for it.

It’s like asking the referee to let your own robot call the plays. The referee says fine, but only if everyone gets to see how you built the robot.

What it means for you: If you want AI to do sensitive review work in a fight, plan on showing how it was trained, tested and checked. “Trust us” isn’t going to cut it.

👉 Read more (New York Post) · Daily Journal

🔁 AI Searches Without the Experts? Do It Over

In Lighthouse Healthcare Center v. National Fire & Marine Insurance, a federal court in California had ordered both sides to bring a litigation support specialist into their talks about how to search two company databases. According to EDRM’s write-up, the specialists only took part in the first meeting. The defendant told the other side its AI searches had problems, but those problems were never worked through with the experts, and the defendant built its own search plan without the other side’s input.

The result, in EDRM writer Michael Berman’s words: “Defendant should have been more transparent.” The court ordered a do-over in which the parties must meet again and bring in an independent, neutral search expert.

It’s like two neighbors agreeing to have a surveyor mark the property line, then one of them marking it himself with a phone app. Now everybody’s paying for a new surveyor.

What it means for you: AI search is fine, but it has to be discussed in the open. If a court tells you to involve experts, keep them involved.

👉 Read more (EDRM)

🧑‍🔬 Expert Witnesses and Their AI Prompts

Lawyers at Haug Partners rounded up how courts are treating expert witnesses who use AI. When experts handed in reports with made-up, AI-generated citations, courts threw the work out. In one case, the court said the fake citations “shatter[ed]” the expert’s credibility. An expert who used ChatGPT only after the report was written was allowed to testify.

The bigger news is about prompts, the instructions people type into AI. In one federal case against Shell, a court said an expert’s AI prompts were part of the expert’s method and had to be handed over, even though the parties had an agreement limiting expert discovery. In a Texas case against 3M, more than 300 pages of prompts came out in the middle of a deposition, and the jury later awarded $61.5 million.

Think of prompts as the expert’s recipe card. If the cake ends up in court, the other side gets to read the recipe.

What it means for you: Ask your experts whether they use AI and how. Put clear terms about AI prompts in your written agreements with the other side.

👉 Read more (Haug Partners via JD Supra)

📜 California’s New AI Law Reaches the Review Room

California’s SB 574 takes effect Jan. 1, 2027. It requires lawyers to use their own judgment with AI, protect confidential information, check accuracy, and personally verify every citation they file. Arnold & Porter’s eDiscovery team says it could reach further than people think. Citations to the case record pulled from AI summaries may count, and AI-prepared discovery responses and privilege logs still need to be checked, even though discovery papers get an exception from the citation rule itself.

A separate case, Experian v. Stein Saks, shows how quickly review budgets can blow up. An insurer refused to cover litigation costs just as review of more than 700,000 documents was starting. The court said the late privilege claims weren’t waived, but it ordered a proper, document-by-document privilege log.

What it means for you: If you work in California, check AI-drafted discovery work like you’d check a new hire’s. And confirm your AI vendor’s access controls before loading client data.

👉 Read more (Arnold & Porter) · Experian v. Stein Saks (EDRM)

🔐 A Backup Breach and Disappearing Logs

Microsoft said a China-aligned hacking group it calls Storm-0940 broke into ConnectWise Cloud Backup, a service that backs up Microsoft 365 accounts. According to ComplexDiscovery, some of the attacks were run by AI agents and reached high-level cloud access within three hours, and legal is one of the industries Microsoft listed as targeted. The catch: on Microsoft’s free Entra tier, sign-in and audit logs are kept for only seven days, so the records you’d need to investigate may already be gone.

A separate report from security company Straiker describes a criminal group that ran a team of AI models, assigning different jobs to different AIs, to go after more than 400 Latin American financial targets.

What it means for you: Ask your IT team or provider how long your Microsoft 365 logs are kept and whether your backup vendor was affected.

👉 Read more (ComplexDiscovery) · Straiker’s report

🔌 Tools Plugging Into Each Other

LexisNexis’s Protégé AI now works inside EvenUp, a tool personal injury lawyers use, so they can ask legal research questions using facts from their own case files. Answers come with linked citations and Shepard’s signals. Candle AI joined iManage’s partner program to connect email with the firm’s document system. And at NetDocuments’ Inspire conference, the company put its “Legal Context Graph” at the center of its pitch, while Above the Law warned that “AI is going to cost more as the industry shifts from subscription pricing to consumption pricing.”

What it means for you: When tools share information, ask what’s being shared, and whether there’s a record of it.

👉 Read more (LawSites) · NetDocuments Inspire (Above the Law)

🏛️ Rules of the Road

• Show the training, show the testing. A federal judge said AI privilege review would only be considered if a model’s training and testing were fully disclosed. New York Post
• Don’t bill for the time AI saved. Guidance from England’s Law Society says firms that bill by the hour “should not bill for the time that a task would have taken had AI not been used.” Legal Futures
• Verify it yourself. California’s SB 574 requires lawyers to personally check every citation they file, starting Jan. 1, 2027. Arnold & Porter

🧭 Quick Hits for the Desk Drawer

• The billable hour hangs on. 8am’s data on more than 60,000 firms found 47.7% bill entirely by the hour, up from 46.6%. LawSites
• AI found the bugs first. Researchers used OpenAI models to find two critical flaws in PHP, which runs a big share of the web, in a matter of days. eDiscovery Today
• A detector that passed one test. eDiscovery Today ran the IdeaLens AI detector on a post the author wrote himself. It said 1.6% likely AI, while an older detector had said 92%. eDiscovery Today
• An AI helper that watches your apps. Anthropic previewed Claude Duet, an AI agent that sends daily briefs and “checks with you before it acts on your behalf.” CNBC
• Privacy requests turned lawsuits. EDRM and Exterro host a free webinar Oct. 28 at noon Central on when data access requests become eDiscovery projects. eDiscovery Today
• Level up slowly. An Above the Law column says to learn AI the way you’d level up a Dungeons & Dragons character, one encounter at a time. Above the Law

✅ Your Front-Porch eDiscovery AI Checklist

• ✅ Write down how any AI review tool was trained, tested and checked.
• ✅ Keep court-required tech experts in every search meeting.
• ✅ Tell the other side about AI search problems, and work them out together.
• ✅ Ask expert witnesses whether they use AI, and save their prompts.
• ✅ Put AI-prompt terms in your written expert agreements.
• ✅ In California, verify AI-drafted discovery responses and privilege logs.
• ✅ Check how long your Microsoft 365 logs are kept.
• ✅ Bill for real time worked, not time the AI saved.

Practical rule: If AI helped you get the answer, be ready to show the judge how.

— Dunlap.Today

📚 Plain-English AI Glossary

• eDiscovery: Finding, saving and handing over electronic information (emails, chats, files) for a lawsuit or investigation.
• Privilege review: Going through documents to find the ones protected as private lawyer-client communications, so they aren’t handed over.
• Privilege log: A list describing each document held back as privileged and why.
• Special master: A neutral person the court appoints to oversee part of a case, like checking a review.
• Meet and confer: Required talks where both sides try to agree on things like how to search for documents.
• ESI: Electronically stored information, meaning emails, files, chats and data.
• Prompt: The instructions or question a person types into an AI tool.
• AI agent: AI that can plan and carry out several steps on its own, not just answer one question.
• Audit log: A record of who signed in and what they did, and when.
• Connector (MCP): A standard plug that lets one AI tool ask another system for information.

❓ Frequently Asked Questions

Did the judge ban AI from document review?

No. She rejected Meta’s request for this particular privilege re-review and said she’d only consider it if a model’s training and testing were fully disclosed. AI review is still used in many cases, often by agreement.

Why did the court order a do-over in the Lighthouse case?

The court had told both sides to involve tech experts when planning searches. They mostly didn’t, AI search problems weren’t worked through together, and one side built its own plan. So the court brought in a neutral expert.

Can the other side really read my expert’s AI prompts?

Increasingly, yes. At least one court treated an expert’s prompts as part of their method. Ask your lawyer to address it in your written agreements.

Does California’s SB 574 apply to me now?

It takes effect Jan. 1, 2027, and applies to lawyers practicing in California. Now’s a good time to set up checking routines.

🔗 Other Resources

• Daily Journal: Meta agrees to review 170,000 privilege claims
• Legal Futures: Law Society says you may sometimes need client consent to use AI
• LawSites: Candle AI joins the iManage partner program
• EDRM on X: ComplexDiscovery eDiscovery Pricing Survey, open through Nov. 20

🌄 No Need to Worry, Just Stay in the Know

None of this means AI is off-limits in discovery. Courts are just asking the same question that teacher asked: how did you get there? The people who keep good notes, keep their experts involved and check the AI’s work will be the ones who get to keep using it.

We’ll keep watching what matters for eDiscovery providers and their clients. Read Dunlap.Today and sign up for our newsletter for plain-English updates.