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

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., 4^th Dist. Ca. Nov 4, 2026)(unpublished).[1]

While the court addressed several other issues, as to the expert’s use of A.I., it wrote: “We decline to disturb the court’s credibility finding that the expert’s use of AI had no impact on the independence of the expert opinion….”

Ms. Gillespie was charged with grand theft. She provided expert evidence of her mental health issues, and filed a motion for pretrial mental health diversion.  Under California law, the expert evidence raised a presumption that mental health was a significant causative factor to consider on the diversion motion.

The People opposed diversion.  They asserted that there was clear and convincing evidence that the mental health issues were not a significant causative factor.  “To rebut the presumption of a causal nexus, the prosecution relied on a report from its own mental health expert, Dr. Nicole Friedman.”

And, that is where A.I. came into the picture.  The court wrote:

Friedman’s assessment of Gillespie was based on the arrest report, the complaint and charging documents, Gillespie’s motion for diversion and supplemental filing, including both of [defense expert] Vickerman’s expert opinion reports, and the People’s opposition. Friedman did not interview Gillespie, but she reviewed two short surveillance videos of the alleged offense (each 30 seconds or less in length). Important here, Friedman disclosed in her report that she had used AI to help generate the report. Friedman stated, “AI was used to help organize timeline and symptoms and[/]or behaviors to demonstrate the nexus.” [emphasis added].

Dr. Friedman agreed with, and disagreed with, some of Ms. Gillespie’s expert’s opinions.  But, Dr. Friedman’s report contained an error: “Friedman incorrectly reported that, prior to the date of the offense, Gillespie’s medical records contained ‘[n]o documentation’ of an acute psychiatric hospitalization, manic episode, psychotic symptoms, or severe functional decompensation.”

At the diversion hearing, “Defense counsel asserted [that Dr. Friedman’s] report was unreliable and should be excluded because the prosecution would be unable to establish that Friedman’s reliance on AI satisfied the Kelly/Frye test for the admission of evidence that is based upon the application of a new scientific technique.”

The [trial] court rejected the contention, stating that use of AI was “very common” and “just because something involves AI” does not mean “that automatically, there’s a problem with it.” The court interpreted Friedman’s report to disclose that AI was used to develop a timeline and said the defense would be entitled at the hearing to probe further about Friedman’s use of AI to arrive at her opinion.

Dr. Friedman testified and, in pertinent part:

As for her use of AI, Friedman testified she uses ChatGPT under a business associate agreement, meaning it prevents the inputs from becoming part of the training model. She uses ChatGPT as “a thought partner … particularly, in regard to the legal questions at hand.” She also uses it to “rewrite” and “polish” her sentences. She could not provide her ChatGPT inputs for the case because she deletes “everything” after she uses it. When asked specifically if she prompted ChatGPT with any questions about the potential nexus between Gillespie’s diagnoses and the charged offense, Friedman remembered that she asked, “What would be, maybe, considered nexus or causal effect.” But she repeatedly denied using ChatGPT “to come up with an opinion.” She explained, “I used it more as an organizational tool; not as something to opine or come up with my conclusions.”

On cross-examination, defense counsel pointed to the statement in Friedman’s written report that Gillespie’s medical records contained “‘no documentation of acute psychiatric hospitalization.’” Friedman admitted the statement was inaccurate and said that she left the information out of the notes she provided to ChatGPT, but considered it when formulating her opinion. To further impeach Friedman’s testimony, the defense presented testimony from another expert, Kevin Diosdado, who ran Friedman’s report through an AI detection program. Diosdado testified the AI program detected the nexus section of Friedman’s report was “98 percent AI.”

At the end of the hearing, the trial court found Gillespie ineligible for diversion. The court explained it had relied upon and was persuaded by Friedman’s expert opinion….  As for Friedman’s use of AI, the court found, “there was nothing in … Friedman’s testimony that led me to believe that the opinions that she gave and the reasons she gave for those opinions were impacted in any way by the AI.”  [emphasis added].

On appellate review, Ms. Gillespie argued that the testimony showed that Dr. Friedman used AI to formulate her opinions; that use was a novel technique; and, it lacked sufficient indicia of reliability.

The appellate court wrote: “Although we share Gillespie’s concerns about Friedman’s reported use of AI, we must defer to the trial court’s credibility findings and consequently find no error on that basis.”  On an abuse of discretion standard, the court found no abuse:  “When deciding whether expert testimony is inadmissible under these two provisions, a trial court ‘must simply determine whether the matter relied on can provide a reasonable basis for the opinion or whether that opinion is based on a leap of logic or conjecture. The court does not resolve scientific controversies.’”  It added:

We agree with Gillespie in principle that the use of AI by a forensic psychologist to generate an expert opinion, at present, would constitute a novel scientific technique. Any such opinion, therefore, would be inadmissible absent an adequate foundation demonstrating the reliability of the psychologist’s methodology, including the “general acceptance of the new technique in the relevant scientific community.” … But Friedman testified repeatedly that she used ChatGPT “as an organization tool,” and did not use it to generate her opinion. She testified that ChatGPT “was not something that [she] used as creating [or] relying on to come up with an opinion.” She explained, “My opinion relies on the totality of the records that I reviewed, which includes … [defense expert] Vickerman’s report, that includes the video, that includes understanding what the Penal Code [says]. So it’s not just what … what ChatGPT might have told me.” [emphasis in original].

The appellate court deferred to the trial court’s finding that the expert opinion was not based on a novel, unestablished scientific technique: “This is because, according to her testimony, she used it to help organize and focus her thoughts but then independently analyzed the evidence and drew her own conclusions.”

It did so despite “troubling” evidence:

Pointing to evidence from which a reasonable factfinder might question Friedman’s representation that she conducted an independent analysis, Gillespie’s argues that Friedman delegated critical components of her forensic analysis to ChatGPT in a way that compromised her independence. She contends this evidence “established that [Friedman’s] use of ChatGPT extended well beyond clerical organization and into substantive legal and psychological analysis central to the court’s eligibility determination.” Specifically, Friedman testified she used ChatGPT as a “thought partner.” She posed legal questions about the Penal Code to help her understand the relevant provisions, and she asked it to answer the central question critical to her analysis: “What would be, maybe, considered nexus or causal effect.” Despite claiming an independent review, she also failed to notice that her written report, drafted with ChatGPT’s assistance, inaccurately stated that Gillespie’s medical records contained “[n]o documentation” of an acute psychiatric hospitalization, manic episode, psychotic symptoms, or severe functional decompensation—a glaring error. We agree this evidence is troubling and of the type that might lead a reasonable trier of fact to conclude that Friedman’s opinion was not independent, as she claimed, but rather based on an unproven and unreliable delegation of her analysis to ChatGPT. [emphasis added].

That “troubling” argument failed under the standard of review:

But the trial court drew a different conclusion. When asked what she meant by a “thought partner,” Friedman analogized her use of ChatGPT to using a colleague as a sounding board but ultimately coming to her own conclusion as to her opinion. The trial court credited this account and additionally analogized her use of ChatGPT to the use of an intern and research attorneys by a partner at a law firm. The court explained, “there was nothing in Dr. Friedman’s testimony that led me to believe that the opinions that she gave and the reasons she gave for those opinions were impacted in any way by … AI. So there are times when you think through something[;] she called it a thought partner. But it’s a matter of processing…. [I]n the end, you step back and make your own decision. And it seemed to me that’s exactly what she was doing. And she supported it in her testimony, and she supported it in her report too.” [emphasis added].

As such, there was no abuse of discretion.  The trial court did not err in finding that Ms. Gillespie was ineligible for diversion.

It is noteworthy that the expert destroyed her prompts and that was not apparently an issue.  Are an Expert’s A.I. Prompts Discoverable?  Maybe. – E-Discovery LLC (Oct. 12, 2026).

For more information on use of AI by experts, please visit:

• Expert’s Use of GenAI Was Proper on Facts Presented – E-Discovery LLC (Oct. 8, 2026);
• Opposing Expert’s Allegedly Inaccurate Description of Use of A.I. Did Not Justify Modification of Scheduling Order on Facts Presented – E-Discovery LLC (Sep. 15, 2026);
• “‘Show How 3M Is 0% at Fault:’ Expert Witness Used ChatGPT to Write Report Defending Company in Deadly Explosion Lawsuit” – E-Discovery LLC (Aug. 26, 2026)

____

[1] California Rules of Court, Rule 8.1115, restricts citation of unpublished opinions in California courts.  Maryland has, in part, eliminated the fiction of disappearing judicial ink.  Elimination of Judicial Disappearing Ink in Maryland – E-Discovery LLC (May 6, 2023); Maryland’s New and Improved Unreported Opinion Rule Does Not Go Far Enough – E-Discovery LLC (Jun. 1, 2024).