Medical records are often treated as the objective ground truth of a patient’s health. A new study by an epidemiologist and a sociologist suggests that is a comforting illusion, and it matters more now that hospitals are building AI on top of those records. The researchers, Annemarie G. Hirsch and Jennifer Silva, paired electronic medical records with detailed interviews with 87 women about their pain, illness, work, family, poverty, violence, and encounters with the health care system. The gap between what the women said and what the charts contained is the part worth pausing on.
The distance between a life and a chart
Start with the numbers. Fifty-three of the 87 women described an encounter that broke their trust in a health system. Only seven of them had records that documented any mistrust. Thirty-two women reported being misdiagnosed, yet misdiagnosis appeared in the records of only eight of them. Thirty women said clinicians treated them with disrespect, and that appeared nowhere in their records at all.
These findings sit alongside a broader survey. The Gallup-Pivotal Women’s Health Survey, reported in September, found that 34% of U.S. women strongly or somewhat agreed that a provider had minimized or dismissed their symptoms in the past five years. Thirty-three percent reported being misdiagnosed or unable to get a diagnosis, compared with 21% of men.
None of this means the records are falsified. It means they are incomplete in a patterned way. A record captures what a clinician asked, what the clinician chose to write down, and what the patient felt safe enough to say. Delays, dismissals, and damaged trust tend to fall outside all three.
What one chart shows and leaves out
The authors illustrate this with a woman they call Kate, a pseudonym. For more than a decade she lived with debilitating pelvic pain, and her doctors repeatedly tied her symptoms to her weight. Her notes carried that assumption forward, with lines suggesting she could not control what she ate. After bariatric surgery and significant weight loss, the tone of the notes turned optimistic, even celebrating a half marathon. Her symptoms persisted, so the record pivoted again, now attributing them to mental factors.
The turning point came when she attended an appointment with her new husband. The provider documented his concerns about painful intercourse and noted that he had very good questions. For the first time in ten years, doctors conducted the exams needed to diagnose her with endometriosis.
Here is what the record got right and what it missed. It accurately logged years of conversations about weight and mental health. It did not log the cost of an undiagnosed disease, the frustration of not being believed, or the slow erosion of trust. In Kate’s words, every symptom was explained by being overweight, and no one was hearing what she was saying.
This is the detail most discussions of clinical AI skip. A chart is a narrative, and narratives carry framing. Language that reinforces a stereotype can follow a patient from one encounter to the next, shaping how later clinicians read her symptoms and judge her credibility.
Why this becomes an AI problem
Clinics and hospitals across the country are adopting AI scribes, which transcribe a recorded conversation between doctor and patient and produce a clinical note. Other tools search and synthesize records to support diagnosis and treatment planning, some with real success.
The catch is simple: these systems are only as good as the data they read. If a woman’s symptoms are filtered through assumptions about her weight or mental health, and the true story of her pain never makes it into the record, an AI system can learn that framing and reproduce it. The bias does not need to be programmed in. It arrives quietly through the training and input data.
There is also a subtler layer. A scribe that transcribes a conversation can only capture what is said aloud. If a patient holds back because she fears judgment, or feels pressure to stay pleasant and agreeable so she will be taken seriously, the omission travels straight into the note. The authors point out that this pressure falls particularly hard on women. Better transcription technology cannot fix a conversation in which the most important part was never spoken.
What a better record would take
The authors do not argue against AI. They argue that its potential in women’s health depends on records that tell the whole story, and they name three conditions. Health care settings must be safe spaces where women can tell their stories. Clinicians must listen even when an account does not fit clinical expectations. And clinicians must be able to document delayed or incorrect diagnoses without fear of being penalized.
That third condition is a human and institutional one, and it is the hardest. Honest records of past mistakes are the raw material for learning from them. If recording an error carries a penalty, errors stay invisible, and invisible errors are exactly what a pattern-finding system cannot see.
With fuller data, the authors suggest AI could do more than replicate the biases of the past. It could help surface patterns of missed diagnoses, reduce disparities in care, and make sure more women are heard. In this framing, the technology amplifies the people in the room: it can extend a clinician’s attention across thousands of cases, but only if clinicians and patients first give it an honest story to work with.
In short
Medical records are narratives shaped by clinical questions, assumptions, and patient trust, which is why 53 women described broken trust while only 7 charts recorded it. AI tools that read these records inherit whatever is missing from them, including the framing that led to years of dismissed symptoms. The remedy starts with people: safer conversations, genuine listening, and the freedom to document diagnostic mistakes. With a more complete record, AI could help clinicians spot missed diagnoses and hear more of the women they serve.
Based on reporting from STAT News.