Insights / The Value of Nurse-Verified AI
What Happens When AI Misinterprets a Medical Record?
By Breen LNC.
Short answer: A misread entry can travel. It can move from a chronology into a demand letter or a deposition outline, where the defense or an expert may be the first to catch it. The fix is to check each entry against its page before any other work is built on it.
How errors enter the record
Some errors come from the record itself. A systematic review of copy and paste in electronic charts identified note bloat, internal inconsistency, error propagation and documentation in the wrong patient’s chart among the risks, while noting that evidence of direct patient harm is sparse. An AI tool reading that chart inherits those problems and can make a stale line look current.
Others come from the software. Researchers studying a speech recognition system found that about 1% of transcriptions contained entire invented phrases, and 38% of those included explicit harms such as false associations or implied authority. That was speech-to-text on more than 13,000 clips from a research collection of speakers with and without aphasia, and hallucinations were more frequent among speakers with aphasia. A researcher quoted by Cornell says the model has since improved, and the study doesn’t measure any clinical product. It does show that software can add words nobody said.
How one error spreads
Consider a hypothetical. A chronology entry marks a medication as given when the chart shows it was held. The attorney builds a theory of care around it. The demand quotes it. At deposition, the nurse who charted the dose corrects it under oath.
We use a made-up example because we won’t invent case histories. Real sanctions cases show how the pattern plays out with legal citations. In Mata v. Avianca, the order found the lawyers kept standing by fake opinions after the court questioned them. Washington’s advisory opinion describes Kohls v. Ellison, where a court struck an expert declaration that cited non-existent academic articles generated by an AI tool.
What it costs
The Mata court listed the harms from fake opinions, including wasted time and money and the possible loss of the client’s authentic arguments. In a medical case you could add a lost chance to find what the record really showed. Credibility is hard to measure, and it tends to be hard to win back.
Stop it at the first step
Errors are cheapest to catch early, before anything has been built on them. Review each entry against its source before it reaches the attorney, and require a page cite so that anyone downstream can check it too. That’s the checkpoint Breen builds in.
Frequently asked questions
Can AI add information that is not in the source?
Yes. Research on a speech recognition system found about 1% of transcriptions contained entirely invented phrases. That study covered audio transcription, not record summaries, and the model has since been updated.
What is the best way to catch AI errors in a medical record review?
Check each entry against its cited page before other work relies on it. Errors caught at that stage have not yet reached a demand or an expert.
Sources
- Tsou et al., Safe Practices for Copy and Paste in the EHR, Applied Clinical Informatics (2017)
- Koenecke et al., Careless Whisper: Speech-to-Text Hallucination Harms, ACM FAccT (2024)
- Cornell Chronicle, AI speech-to-text can hallucinate violent language (June 11, 2024)
- Mata v. Avianca, Inc., No. 22-cv-1461, Opinion and Order on Sanctions (S.D.N.Y. June 22, 2023)
- Washington State Bar Association, Advisory Opinion 2025-05, Artificial Intelligence-Enabled Tools in Law Practice (2025)
General information for legal professionals. This is not legal or medical advice.