Home › News › Even With an Error, Expert’s “Troubling” Use of AI as a “Thought Partner” Did Not Undermine Expert Opinion

Even With an Error, Expert’s “Troubling” Use of AI as a “Thought Partner” Did Not Undermine Expert Opinion

Industry newsAi in litigation

An AI-Assisted Expert Survived Challenge in California, but the Deleted Prompts Are the Real Warning

On November 4, 2026, California's Fourth District Court of Appeal denied a defendant's petition for interlocutory review in Gillespie v. The Superior Court of San Diego County, 2026 WL 3333170, an unpublished ruling. Ms. Gillespie faces a grand theft charge and sought pretrial mental health diversion. The prosecution rebutted her expert with Dr. Nicole Friedman, who disclosed that AI helped "organize timeline and symptoms" in her report. According to Michael D. Berman of E-Discovery LLC, writing for EDRM on November 11, 2026, the appeals court declined to disturb the trial court's finding that the AI use did not affect the independence of her opinion. This note is current as of November 16, 2026.

What did the court actually decide about the expert's AI use?

The court declined to disturb the trial judge's credibility finding that Dr. Friedman's AI use did not affect the independence of her opinion. That is a deferential ruling, not an endorsement of ChatGPT.

The defense argued that the report failed the Kelly/Frye test for new scientific techniques. Per EDRM's account, the trial judge rejected that, said AI use is common, and said involvement alone is not a problem. The defense could probe further at the hearing. Dr. Friedman testified that she uses ChatGPT under a business associate agreement and treats it as a "thought partner," according to Berman's account. She also uses it to rewrite and polish sentences.

Contrary to the reassuring reading, the case does not show that courts have validated AI-assisted analysis. It shows how little a reviewing court will second-guess a trial judge's view of one witness.

Does the error in the report undermine AI-assisted expert work?

Not legally here, but it should worry buyers. Dr. Friedman's report wrongly said Gillespie's earlier records held no documentation of psychiatric hospitalization, manic episode, psychotic symptoms or severe decompensation, yet her opinion stood.

Neither EDRM nor the court says the AI caused the mistake. But she disclosed using AI on the timeline, which is where a missed record would surface. She also testified that she deletes "everything" after use, so no one can test whether the tool dropped the hospitalization history or she did. Nobody can answer that, and that is the problem.

What is the Three-Record Test for AI in expert work?

The Three-Record Test is our shorthand, not the court's. An AI-assisted opinion is defensible only if three records exist: what went in, what came out, and who checked it.

Apply it to Gillespie. The business associate agreement speaks to the input record, because it keeps inputs out of the training model. But a contractual promise about training is not a log. The output record is gone, because she deleted her inputs. The verification record is the error itself: a report with a wrong statement about the medical file. Two of the three records failed, and the opinion survived on the judge's view of her credibility.

What should buyers ask vendors and experts now?

Ask whether every prompt, output and source citation is retained, exportable and tied to a source document, and who controls deletion. If the answer is "the user," you have no record.

Ask who defines "private." A business associate agreement is the vendor's own contractual definition. It says nothing about retention or reproducibility. This is where evidence platforms differ from general chat tools. A system that issue-codes a collection, builds a cited chronology and links each answer to a source exhibit, run single-tenant or in the firm's own AWS account, leaves the record that Gillespie lacked. We have made related points in AI and Legal Work: Evaluating Value, Risk and Responsibility and Authenticating What You Cannot Trust.

Law.com ran a related opinion on November 12, 2026, in which Thomas F. Berner argues that AI output needs careful human review. The text is paywalled, so we draw only on its summary.

Frequently asked questions

Is Gillespie binding precedent?

No. The opinion is unpublished, so treat it as a signal of judicial attitudes, not a rule. Check with counsel before citing it.

Did the court approve ChatGPT for expert reports?

No. It deferred to a credibility finding on independence and did not rule on whether the tool is reliable.

Can an expert delete AI inputs after finishing a report?

Dr. Friedman did, and her opinion survived. The EDRM text we reviewed cuts off mid-testimony, so we cannot say how far cross-examination went.

Sources: EDRM, "Even With an Error, Expert's 'Troubling' Use of AI as a 'Thought Partner' Did Not Undermine Expert Opinion" (Michael D. Berman, November 11, 2026); Law.com, "AI Is Not Superintelligent; It's Not Even Intelligent" (Thomas F. Berner, November 12, 2026).

The original report

The defendant in a criminal case sought interlocutory review of an order permitting expert testimony that was “assisted” by artificial intelligence. Her petition for interlocutory review was denied. Gillespie v. The Superior Ct. of San Diego County, 2026 WL 3333170 (Ct. Appl., 4th Dist. Ca. Nov. 4, 2026)(unpublished).

EDRM

Read it on EDRM

Industry headlines from other publications. Each links to the original reporting on the publisher's own site.