In 2025, a federal judge excluded an expert witness report because the citations inside it were hallucinated. Not misinterpreted. Not contested. Fabricated, the papers did not exist. The expert had used AI assistance in preparing the report. Nobody verified the citations before filing.
The case is Kohls v. Ellison. It is the first landmark precedent in the intersection of AI-assisted expert work and evidentiary admissibility. And it is a warning shot for every expert witness who has used AI in their preparation workflow without building verification into it.
What Happened
An expert report, assisted by AI, contained hallucinated citations. The judge excluded the report. The case outcome changed.
The facts as established in the court record:
- An expert witness submitted a report supporting a causation claim in federal litigation
- The report contained citations to academic studies, real-looking DOIs, author names, journal references
- Opposing counsel challenged the citations. They did not resolve
- The expert could not produce the cited papers because they did not exist
- The judge excluded the report under Daubert, finding the methodology unreliable
The exclusion was not about the expert’s ultimate conclusion. It was about the verifiability of the sources the conclusion rested on. A report that cites non-existent papers as methodological support is a report whose methodology cannot be tested, challenged, or trusted.
The case landed during a period of intense judicial scrutiny of AI-assisted legal work. By early 2026, there were 156 documented lawyer sanctions for AI hallucinations across federal courts (LegalDive, compiled from federal court records). Judges were no longer issuing warnings. They were excluding evidence.
The 156 Sanctions Pattern
This is not one case. It is one visible case.
The 156 documented lawyer sanctions for AI hallucinations tell the broader story. Each one represents an attorney or expert who used AI-generated content, case citations, regulatory references, academic sources, without verifying that the content was real.
The pattern is consistent:
- AI generates text with plausible-looking citations
- The human reviewer recognizes the topic as relevant and accepts the citation
- Opposing counsel checks the citation. It does not exist
- The court sanctions the filing party
The sanctions range from monetary penalties to adverse inference instructions to, in the most serious cases, evidence exclusion. Kohls v. Ellison is the exclusion case, the one where the AI hallucination was not just sanctioned but used to disqualify the entire expert report.
The distinction matters because sanctions are a penalty. Exclusion is a case outcome. An expert who pays a fine still has their report admitted. An expert whose report is excluded has lost the evidentiary foundation for their testimony.
Rule 707: What Comes Next
Proposed Federal Rule of Evidence 707 would make AI disclosure mandatory. It is currently under public comment.
Rule 707, if adopted, would require attorneys and expert witnesses to disclose any use of generative AI in preparing filings and reports. The rule would not ban AI use. It would make it transparent, so that opposing counsel, judges, and juries can evaluate whether AI-assisted work was properly verified.
The rule’s existence is itself a signal. The federal rulemaking process does not create disclosure requirements for tools that are not causing problems at scale. Rule 707 is the institutional response to the 156 sanctions and the cases behind them.
For expert witnesses, Rule 707 changes the calculus. Disclosure is not the risk, unverified output is. An expert who discloses AI use and produces a verification trail showing that every citation was checked, every source confirmed, and every claim aligned to evidence is in a stronger position than an expert who used AI but cannot demonstrate any verification.
The rule does not punish AI use. It punishes unverifiable AI use.
The judge didn’t exclude the report because it was wrong. She excluded it because nobody verified it before it was filed.
What Verification Would Have Caught
The hallucinated citations in the excluded report had a specific failure profile. A verification system would have flagged them before the report was ever filed.
The failure mode was fabrication, the easiest to catch and the most damaging when missed. A verification system running existence checks (CrossRef, DataCite, PubMed resolution) would have identified every non-existent DOI before the report left the expert’s desk. The flags would have read:
Citation [7], DOI: 10.1037/example.2024.001
Status: FABRICATED, DOI does not resolve in any registry
Recommendation: Remove or replace with verified sourceThe expert would have replaced the fabricated citations with real ones. The report would have reached the court with verified sources. The Daubert challenge would have had nothing to stand on.
The Contradiction Register, the deliverable that documents each verification step, would have produced a record showing exactly what was checked, what was found, and what was corrected. Under Rule 707, that record is the disclosure artifact that demonstrates responsible AI use.
The Case That Changes the Market
Kohls v. Ellison is the case that makes verification infrastructure necessary, not optional, for expert witnesses.
Before this case, the risk of AI hallucination in expert reports was a theoretical concern. After this case, it is an exclusionary precedent. Opposing counsel who know about Kohls will check citations in every AI-assisted expert report. The expert who cannot produce verification evidence is in the same position as the expert in Kohls.
The verification infrastructure that catches hallucinated citations, existence checks, polarity analysis, claim-source alignment, is not a premium add-on for cautious experts. It is the minimum standard for any expert using AI in their preparation workflow.
The question for expert witnesses is no longer whether to verify AI-assisted output. It is whether they can afford to file a report without a verification trail that would survive a Kohls-style challenge.
If your next expert report uses AI assistance and does not include a verification trail, opposing counsel will run the check you didn’t. Request a legal intake at axion.activewizards.com/legal-pilot or reach us at axion@arizenai.com.