Important Court Rulings from 2026 Regarding AI: Artificial Intelligence Trends

Courts treat AI output as the signer's problem, so buy for verification, not speed

In a post dated October 2026, Doug Austin of eDiscovery Today relayed one trend from Exterro's whitepaper "Prompts, Privilege, and Preservation in 2026": a non-delegable duty to verify AI output. In Pauliah v. Univ. of Mississippi Med. Center, an attorney was sanctioned over AI-generated deposition summaries containing fabricated verbatim quotes. In Whiting v. City of Athens, the Sixth Circuit imposed $15,000 punitive sanctions per attorney for hallucinated citations in appellate briefs. In American Council of Learned Societies v. NEH, a federal judge would not let an organisation blame ChatGPT for a flawed grant-screening process.

Does using a vendor's AI tool move liability away from the lawyer?

No. In the cases described, the signing attorney or the deploying organisation bore the consequences, regardless of which AI product produced the faulty text or analysis.

The ACLS ruling is the awkward one for suppliers. It concerned grant screening, not document review, but the reasoning travels. Austin reports that once an organisation makes a platform its "chosen instrument," it owns the output's legal consequences. Nothing in the post suggests a vendor contract could shift that duty. At best, a contract allocates cost afterwards.

Treat the Whiting figure as a single data point. $15,000 per attorney, from an appellate court, over citations, is not a tariff. "Steep" is Austin's word, not a benchmark.

Why do the deposition-summary sanctions matter more than the citation cases?

Because the fabricated quotes sat in summaries of the evidentiary record rather than in legal argument, and review platforms that summarise documents and build chronologies do the same kind of work.

A bad citation can be caught by looking the case up. A bad deposition quote means someone opens the transcript and hunts for the passage, which is the labour summaries are bought to avoid. The post does not say how the invented quotes in Pauliah came to light. That detail matters, and nobody in the coverage supplies it.

A verbatim quote is also the most testable claim a tool makes, because the string either appears in the source or it does not. A platform that links each answer to its exhibit helps, but a link only lets a person check. It does not show that the quoted words match the linked page.

What should I ask my e-discovery vendor now?

Ask how the product proves that each quote, summary line and chronology entry matches its source, and what the audit log shows when a reviewer checks or overrides it.

Press on measurement too. If a vendor quotes an accuracy rate, ask who defined it, who measured it, on whose documents, and whether fabricated quotes were scored separately from missed documents. Ask where prompts and privileged material are processed, and whether anything is retained or used for training. That question is sharper when courts treat the tool as yours.

This fits our earlier coverage of legal AI and the adversarial incentive, and of the Broward judge suspended and fined after being duped by a deepfake. Responsibility keeps landing on whoever relies on the output.

Frequently asked questions

Do these rulings bind my firm?

Binding force depends on your jurisdiction. The post identifies an appellate court only for Whiting and gives no dates, so read the opinions rather than the summary.

Does a link to the source document satisfy the duty to verify?

Nothing in the post says so. A link makes checking possible, but the check still has to happen, and ideally be recorded.

How far should I trust this roundup?

Austin discloses that Exterro sponsors his site, and the post covers a single trend, pointing readers to the whitepaper for the rest. I have not seen the whitepaper, so this note does not cover the other trends.

Sources: eDiscovery Today (Doug Austin), drawing on Exterro's whitepaper.