Expert Report Admitted Despite AI Hallucinations in Citations

From retired Third Circuit Judge Thomas Vanaskie (who had also served on the Middle District of Pennsylvania), and who was serving as a court-appointed Special Master in In re: Valsartan Losartan, and Irbesartan Products Liability Litigation; the decision was handed down Sept. 3, but just came up on one of my searches:

Dr. Sawyer’s citation to non-existent sources due to his use of an artificial intelligence tool without adequate verification of the sources generated by the artificial intelligence tool, while perhaps warranting an award of costs in favor the defense and permitting cross examination of Dr. Sawyer during the trial on his failure to verify the sources cited in his report, does not warrant exclusion of his opinions as they are otherwise the product of reliable scientific methodology and are supported by “good grounds,” especially given “the liberal thrust of the Federal Rules of Evidence, the flexible nature of the Daubert inquiry, and the proper roles of the judge and jury in evaluating the ultimate credibility of an expert’s opinion” ….

This was appealed to District Judge Renée Marie Bumb, who decided that the appeal was moot in light of her opinion granting summary judgment to defendant in this case, but “Defendants have preserved their position should Dr. Sawyer’s testimony be presented in another action in this [Multi-District Litigation].” Here’s an excerpt of the plaintiff’s argument in favor of not excluding the expert opinion:

Defendants devote a substantial portion of their brief to mistaken citations in Dr. Sawyer’s report, insinuating that these mistakes render Dr. Sawyer’s entire report invalid. The record shows otherwise. During his May 2, 2025 deposition, Dr. Sawyer forthrightly explained that he used a software tool employing artificial intelligence to assist him in locating scientific articles and toxicology studies for his report. This tool was meant to expedite literature searches for well-established background information on NDMA. While drafting, a handful of references (ten, to be exact) were inadvertently cited incorrectly in Dr. Sawyer’s report. These errors were largely confined to a two-page section of the report summarizing general background facts about NDMA (such as its carcinogenic classification, its genotoxic potential, and common exposure pathways).

Many of the citations at issue, such as those in footnotes 3, 4, 5, 6, 7, and 14, are used only for introductory or background context and are not central to Dr. Sawyer’s core analysis of NDMA or his application of toxicological principles. These references merely provide general scientific context regarding mechanisms of NDMA metabolism, oxidative stress, or DNA repair and have no bearing on the methodologies Dr. Sawyer applied in forming his case-specific opinion. Footnote 8 contains a broken FDA hyperlink, but the referenced announcement clearly exists and is readily accessible. In footnote 66 the correct studies were cited, but minor formatting or author-order errors scrambled the reference. Finally, footnote 109 relates to background information on Bradford Hill criteria and was not used in the causation analysis itself. In sum, none of these minor citation discrepancies affect the substance or reliability of Dr. Sawyer’s opinion.

And here’s an excerpt from the defendants’ reply:

First, far from “forthrightly explain[ing]” that he used AI in writing his report, Dr. Sawyer falsely testified on multiple occasions that the phantom articles he cited exist and that he had reviewed them. (See Sawyer 5/2/2025 Dep. 63:14-25 (Mem. Ex. 5) (asserting that Yuan 2027 “is a real article that [he] reviewed”); see also id. 70:12-71:7 (“I recall reviewing [the Sokolow paper,] and I included the link … which was functional.”).) Only after repeated questioning by defense counsel on the topic did Dr. Sawyer finally admit the citations were false and that they resulted from his use of either Google or AI. (Id. 72:13-17.) To this day, Dr. Sawyer and Plaintiff’s counsel have not identified the particular tool that Dr. Sawyer used to “create” the fake citations, rendering Dr. Sawyer’s dishonesty even more egregious than that in Kohls v. Ellison, No. 24-CV-3754 (LMP/DLM), 2025 WL 66514, at *3 (D. Minn. Jan. 10, 2025). Plaintiff does not address this squarely on-point authority, effectively conceding its applicability.

Second, Plaintiff’s attempt to paint Dr. Sawyer’s citation to 10 fake sources as insubstantial and “peripheral” (Opp’n at 35), is foreclosed by Dr. Sawyer’s testimony. At his deposition, Dr. Sawyer made clear that he used “wording [taken] directly” from the fake sources in drafting sections of his report. (Sawyer 5/2/2025 Dep. 67:22-24, 73:17-20; see also id. 60:10-23 (agreeing some of the language “is a quote from Yuan”).) Moreover, those non-existent sources comprise the bases of Dr. Sawyer’ causation opinions, which is presumably why he originally falsely claimed to have read them when questioned about them at his deposition. As the Kohls court explained, an expert’s “citation to fake, AI-generated sources … shatters his credibility with th[e] Court” and “undermine[s] [the expert’s] competence and credibility[.]” Kohls, 2025 WL 66514, at *4-5. That is precisely what happened here, which should be dispositive.