Expert Reports That Cite Sources That Do Not Exist

The outcomes in this record range from a report struck before trial to an opinion left entirely in place. Tribunals have excluded declarations, withdrawn reports in mid-trial, set a court-appointed witness's fee at zero, found an opinion not credible, reduced its weight, and in several matters left the defect to cross-examination. The result has tracked the kind of defect, how central the material was, and whether the author could account for it, rather than the mere use of a system.

Eighteen matters and their outcomes

The table collects eighteen matters in which a report, declaration or opinion offered as evidence rested on a source that could not be found, or on output nobody could verify. It records who wrote the defective material, because that differs: a retained witness, a court appointee, a treating clinician, a party, or counsel formatting a reference.

Fabricated, misattributed, unverifiable or unreproducible

Four different defects run through these matters, and tribunals have treated them differently.

• A fabricated citation points to a source that does not exist at all. The two articles in Kohls, the Yellow Book sections in Riddle Aggregates and the so-called Aharoni precedent in Avergil are examples.
• A misattributed citation or quotation concerns a real source described wrongly: the wrong authors on a real article in Kohls, the wrong title and authors in Concord, a ruling assigned to the wrong court in Avergil. Misattribution can pass a check that asks only whether something by that name exists.
• Unverifiable statutory or guideline text is a quotation of law or policy that does not match the official text. Avergil's report gave versions of three sections of the Payment Services Law that did not match the enacted text, and the keratoconus letter relied on a Veterans Affairs Canada guideline that the panel found does not exist.
• An analysis that cannot be reproduced is a figure or conclusion no one can retrace. In Weber the prompts were gone; in Celsius no standards controlled the system that wrote the report.

The first three are defects in what the book calls the citation apparatus; the fourth is a defect in the opinion itself. The Kohls order and the Weber decision illustrate the two sides.

Struck, excluded or withdrawn

Struck or excluded before the evidence was heard

In Avergil the judge ran her own check, found four kinds of defect, and struck the banking report under the civil procedure regulations without waiting for cross-examination. She held that the fictitious sources "strip the opinion of any professional value" (the record's translation from the Hebrew). In Kohls the court did not consider the declaration and refused the corrected version, writing that the citations shattered the declarant's credibility even though it found his explanation helpful and plausible. LeDoux borrowed the shattered-credibility language when it excluded a third plaintiff's witness for multiple hallucinated citations.

Celsius excluded the whole report, but not the witness: his live testimony was admitted to the extent it shed light on the questions. Concord was narrower still. Only paragraph 9 was struck, because the source was real and a person had found it.

Withdrawn in mid-trial

Riddle Aggregates produced no order. On voir dire the Commissioner's appraiser could not explain clearly what he had used, and the judge said she had "grave concerns about this report." After a recess the Commissioner withdrew both reports and called a different rebuttal witness.

Fee set at zero, weight reduced, credibility rejected

Fee set at zero

German procedure gave the Darmstadt court a remedy with no direct US counterpart. It set the court-appointed medical witness's fee at zero against 2,374.50 euros claimed. The full text shows AI was one of three grounds: the author never confirmed that he prepared the report himself, the report was unusable because the claimant was never examined, and the hours billed were disproportionate.

Found not credible or given reduced weight

The keratoconus panel tested each reference by link, issue, page, author and title, found the guideline and articles absent, and held: "The Panel finds that this medical opinion is not credible." The author was a treating clinician writing a supporting letter, not a retained litigation witness, yet the letter was the only medical evidence on causation, and the claim failed without it. Butler reduced the weight of AI-assisted reports where the practitioner could not account for the phrase the funding decision turned on. Weber rejected a damages analysis the witness could not retrace. The Rouen court did not exclude a party's Copilot analysis; it simply declined to count it as professional evidence.

Left to cross-examination, permitted or corrected

Left to cross-examination or permitted

In the Social Media MDL the court declined to exclude Osborne's opinions and said the defendants could raise the citations on cross-examination, without deciding whether the articles exist. The Valsartan Special Master found that the citation problem might warrant costs and cross-examination, but not exclusion of opinions otherwise based on reliable methodology. In Bard the challenger conceded it had found no fully hallucinated citation in the general report, and the court wrote: "Mere suspicion is not enough to exclude El-Ghannam's opinions." Ferlito allowed a witness who consulted ChatGPT only after his report was written. In Watson Grinding, as reported, the witness was examined on his chat logs and no ruling followed.

Two Israeli courts addressed what may be put to a court-appointed witness. Biton required the underlying articles rather than ChatGPT's summaries; Plonit kept an AI-processed summary away from the witness altogether.

Corrected by counsel

Correction after the defect was found has not produced a uniform result. Kohls refused an amended declaration. In Osborne the plaintiffs said the citations had been corrected, and the matter went to cross-examination. In Riddle an errata sheet fixed some citations before trial, but others surfaced in court. In Concord counsel explained that its own formatting step, not the sampling witness, introduced the error. In LeDoux the error also came from counsel, through a citation table two witnesses signed without catching.

What preceded the harsher outcomes in these matters

These are observations about eighteen matters, not rules, and outcomes differ on similar facts. Three features recur in the harsher results.

• No account from the author. Avergil counted against the party that its witness never answered the allegations. In Riddle the judge was not clear what the witness used, and in Weber the prompts could not be recalled. An explanation did not always help, however: Kohls excluded a declaration despite one.
• Repetition. Avergil found four kinds of defect, Holguin's report in LeDoux had multiple hallucinated citations, and in Riddle problems kept surfacing after an errata sheet.
• Centrality. The keratoconus letter was the only causation evidence; the Riddle citations supplied the standard against which the opposing appraisal was measured; the Celsius report was written by a system at the witness's instruction.

The milder results share the opposite features: a real source found by a person (Concord), a disputed defect said to be formatting (Osborne), a methodology the tribunal found reliable on its own (Valsartan), or no fabricated citation shown at all (Bard).

Keeping nonexistent sources out of a report

The obligations that would have prevented each defect are set out in the Standard. Under Clause 3, independent authority, every citation, quotation and statutory passage is confirmed against a source that exists independently of the system. The Avergil entry draws the practical consequence: quote a statute from the official text and keep a copy. Under Clause 4, checkability before adoption, the author names what output would be checked against before adopting it; a book quoted without a page reference cannot be checked by anyone. Clause 5, reproducible analysis, requires a figure to come from a method that returns the same result when repeated, which is what Weber lacked. Clause 2, non-delegation, is the obligation Celsius and Darmstadt turned on.

Clause 7, contemporaneous record, records verification as an act, naming the source consulted. The verification checklist treats a citation as checked only when the source has been opened, and a source that cannot be located as a failure of the assertion: it comes out of the report. Material supplied by counsel is no exception: in Concord and LeDoux the signed document carried every reference, whoever formatted it.

When a challenge arrives, the record suggests answering it in the author's own name, from the log. That is a description of these matters, not advice about any jurisdiction.

Matters this guide relies on

• Avergil v. Yaakov October 8, 2026 · Expert witnesses and AI Primary document read
• Watson Grinding explosion trial (3M expert) August 2026 · Expert witnesses and AI Reported; document not located
• Bard Port Catheter MDL (El-Ghannam ChatGPT challenge) August 12, 2026 · Expert witnesses and AI Primary document read
• LeDoux v. Outliers, Inc. July 24, 2026 · Expert witnesses and AI Primary document read
• Keratoconus Entitlement Review, 2026 CanLII 107034 April 15, 2026 · Expert witnesses and AI Primary document read
• Riddle Aggregates (Tax Court rebuttal expert) March 18, 2026 · Expert witnesses and AI Primary document read
• Social Media MDL (Osborne AI citation tool) February 17, 2026 · Expert witnesses and AI Primary document read
• Cour d'appel de Rouen, RG 25/01677 (Copilot analysis) February 11, 2026 · Expert witnesses and AI Primary document read
• Landgericht Darmstadt expert fee decision November 10, 2025 · Expert witnesses and AI Primary document read
• Biton v. The National Insurance Institute October 16, 2025 · Expert witnesses and AI Primary document read
• In re Valsartan MDL (Dr. William Sawyer) September 3, 2025 · Expert witnesses and AI Primary document read
• Butler and National Disability Insurance Agency August 28, 2025 · Expert witnesses and AI Primary document read
• Concord Music Group v. Anthropic (Chen declaration) May 23, 2025 · Expert witnesses and AI Primary document read
• Ferlito v. Harbor Freight Tools USA April 23, 2025 · Expert witnesses and AI Primary document read
• Kohls v. Ellison January 10, 2025 · Expert witnesses and AI Primary document read
• Plonit v. Clal Insurance (DigitalOwl summary) December 9, 2024 · Expert witnesses and AI Primary document read
• Matter of Weber October 10, 2024 · Expert witnesses and AI Primary document read
• In re Celsius Network LLC (Faraj report) November 9, 2023 · Expert witnesses and AI Primary document read

Questions