Nader Karayanni
TL;DR: Medical experts have always been vulnerable to one line of attack: that the records reviewed were incomplete or selectively chosen.
AI amplifies that attack. And introduces a second: that something in the report was never verified against a source. Both attacks land on the same question: "was your record review complete and unbiased?"
In this blog, we outline the credibility risks specific to medical expert work, what a healthcare case in Utah revealed about unverified AI output, and how patient data obligations fit in.
Key takeaways
Claims managers sometimes send examiners an incomplete set of records. The physician then gets challenged over a review they never controlled.
AI amplifies that exposure. A tool that reads only part of the file, or summarizes selectively, makes an incomplete review look thorough.
A False Claims Act case involving healthcare billing was dismissed with prejudice after an expert's report was found to contain fabricated citations produced with AI.
When using AI for medical record review, use a purpose-built system that is secure, private, and performs a complete and unbiased review of the entire record.
Know the ground rules before you start. What expert witnesses need to know about using AI and AI best practices for expert witnesses and their firms cover what applies to you as an expert using AI in litigation.
Both the old credibility attack and the new one are answered by the same thing: a complete review you can evidence.
Medical experts work with the largest record productions in litigation. Personal injury, workers' compensation, and medical malpractice matters routinely generate hundreds or thousands of pages across providers, formats, and years of treatment. That volume is why AI adoption in medical expert work has moved faster than in most other disciplines.
It is also why medical experts carry a specific credibility exposure that engineering and financial experts do not face in the same way.
What is the oldest attack on a medical expert's report?
Incomplete review. And the uncomfortable part is that it often begins before the records reach you.
Research published in Psychological Injury and Law documents the practice directly. In Cherry-Picking Records in Independent Medical Examinations, Schatman and Thoman describe how "one of the ways in which claims managers contribute to potential bias against claimants is through a process of selectively providing examiners with medical records." The authors note that the practice persists despite regulations designed to prevent it, and that it compromises the theoretical independence of the examination.
The consequence for the examining physician is that an opinion can be formed on a record someone else curated, and then challenged as though the curation were the physician's own choice. Being able to state precisely what was provided, what was reviewed, and what was absent is a direct answer to that line of questioning.
AI can make this worse. A tool that processes only part of a production, or that summarizes selectively without showing what it skipped, produces work that looks thorough and is not. The physician signs it either way. This is why completeness and neutrality matter more in a tool used for medical record review than raw speed does.
What happens when AI errors reach a medical expert report?
A healthcare case shows the outcome plainly. In U.S. ex rel. Khoury v. Intermountain Healthcare Inc., No. 2:20-cv-00372 (D. Utah), the court dismissed the False Claims Act action with prejudice to the relator on September 30, 2025. The underlying allegations concerned anesthesiologists billing federal healthcare programs for care provided while using personal electronic devices.
The expert had used an AI tool to combine three drafts of the report and, as Paul Hastings reported, "failed to verify content in certain instances, despite knowing that generative AI may yield hallucinations." The report contained fabricated deposition testimony attributed to a CMS representative, invented quotations from the Medicare Program Integrity Manual and Medicaid documents, and inaccurate titles for government publications.
Every fabricated element was a government healthcare document, meaning every one was verifiable in minutes by anyone who checked. This is the pattern across the AI exclusions to date, and it is covered further in what expert witnesses need to know about using AI.
How do these two problems connect?
They converge on a single question that opposing counsel will ask in some form: what did you review, and how do you know what it said?
The challenge | What answers it |
|---|---|
You reviewed only what the claims manager sent | A record of exactly what was provided, what was reviewed, and what was identified as missing |
Your tool read part of the file and you presented it as a full review | A system that processes the complete production and shows its coverage |
You relied on a summary and never read the underlying records | Every fact in the report traceable to the page it came from |
Your report cites material that does not say what you claim | Independent verification of every citation, figure, and quotation before signing |
You cannot explain how the review was performed | A process you can describe plainly, including the tools involved |
A complete and unbiased review, with each fact cited to its source, addresses the old attack and the new one at the same time. That is the practical case for using AI carefully in medical expert work rather than avoiding it.
What about patient records and data handling?
Medical expert work carries obligations that engineering and financial work does not. Records contain protected health information, and those handling duties follow the data to wherever it is sent, including any AI tool.
The analysis is genuinely nuanced. Whether and how HIPAA applies depends on the role you occupy in the matter, and an examining physician retained for litigation is not necessarily in the same position as a treating provider. HHS guidance on judicial and administrative proceedings addresses disclosures in litigation, and the protective order in your matter may impose its own requirements.
This is a question to settle with retaining counsel before records are uploaded anywhere, not one to resolve by reading a vendor's marketing page. What you should know going in is what the tool retains, whether your material trains its models, and whether a business associate agreement is available where one is appropriate. We cover the wider set of questions to raise at engagement in AI best practices for expert witnesses and their firms.
Does AI use change your Rule 702 exposure?
Using AI does not by itself affect admissibility. The exclusions to date turned on unreliable content reaching a signed report, which is a Rule 702 problem regardless of what produced it.
What has changed is visibility. A federal court has held that an expert's AI prompts are discoverable as part of their methodology, including prompts used only to narrow document sets. For a medical expert working a 900-page production, that means the way records were organized and searched is potentially examinable, alongside the opinion itself.
Frequently asked questions
Can medical experts use AI to review records and build chronologies?
Yes, and it is the most widely endorsed application. Professional guidance in the US and UK supports AI for summarizing records and building chronologies while cautioning against using it to form opinions or answer the questions you were retained to address. Causation and diagnosis remain yours.
Does using AI on medical records create HIPAA problems?
It depends on your role in the matter and the tool involved. Records carry handling obligations that follow the data, so knowing what a tool retains and whether a business associate agreement is available matters before anything is uploaded. Settle this with retaining counsel at engagement.
What is the biggest AI risk for a medical expert specifically?
Unverified citations reaching a signed report. In the Utah False Claims Act matter, the fabrications were quotations from Medicare and Medicaid documents, all of which were checkable. Verifying every citation, figure, and quotation against its source addresses the failure mode courts have actually punished.
How do I show that my record review was complete?
By being able to state what was provided, what was reviewed, and what was missing, and by tying each fact in the report to the page it came from. That documentation answers both the selective-review challenge and questions about how the review was performed.
What AI tool should medical experts use for record review?
newcase.ai is built for this work: purpose-built for litigation records, private by default with zero data retention, HIPAA-aligned through signed business associate agreements with its AI providers, and every fact tied to a page-line citation in the source document so you can check it before you sign.
Complete and unbiased review is the point. A tool that reads the entire production, applies the same treatment to every record in it, and cites each fact to its page lets you answer the question medical experts have always faced about what was reviewed, with evidence rather than recollection.
Today, expert witnesses use newcase.ai across medical and life care work, engineering and accident reconstruction, financial and economic damages, and alcohol liability, dram shop, and premises security matters, where the record runs long and the opinion has to rest on locatable facts. It is benchmarked across 100,000+ pages of depositions and medical records, whether you are building medical chronologies or working to never miss a fact.
The credibility questions in this post predate AI by decades. The tools now available make them easier to answer.
This post discusses court decisions, professional obligations, and data handling requirements. It is not legal advice. Consult counsel regarding your specific engagement.
Sources
Schatman, M.E., & Thoman, J.L. (2014). Cherry-Picking Records in Independent Medical Examinations. Psychological Injury and Law, 7(2), 191-196
U.S. ex rel. Khoury v. Intermountain Healthcare Inc., No. 2:20-cv-00372 (D. Utah), dismissed September 30, 2025, via Paul Hastings
HHS guidance on HIPAA and judicial and administrative proceedings
Federal Rule of Evidence 702, Cornell Legal Information Institute
Conservation Law Foundation, Inc. v. Shell Oil Co., D. Conn., May 18, 2026, via Spencer Fane
