Medical records can look like objective accounts of what happened during a visit. In practice, they are selective records. They reflect what a patient reported, what a clinician asked, how the clinician interpreted the answers, and what made it into the note.
That distinction matters as health systems use artificial intelligence to summarize visits, search records, and support diagnosis and treatment decisions. AI may find patterns in a chart, but it cannot reliably restore important context that was never documented.
A recent study by epidemiologist Annemarie G. Hirsch and sociologist Jennifer Silva offers a close look at that gap. The researchers compared medical records with detailed interviews with 87 women about their health, work, family lives, and experiences with care. In their interviews, 53 women described an encounter that damaged their trust in a health system. Only seven had mistrust documented in their records. Thirty-two women reported being misdiagnosed, while misdiagnosis appeared in the records of eight. Thirty women said clinicians treated them disrespectfully, and none of those experiences appeared in the records.
These findings do not establish how often such gaps occur among all patients. They do show how much a medical record can leave out, even when it accurately records some of what was said.
When the chart preserves the explanation but not the unresolved problem
The authors describe “Kate,” a pseudonym for one participant who lived with severe pelvic pain for more than a decade. Her clinicians repeatedly connected her symptoms to her weight. After bariatric surgery and significant weight loss, her pain continued, but the records began to link her symptoms to mental health.
The turning point came when Kate attended an appointment with her new husband. The clinician documented his questions about painful intercourse. Kate was then examined for endometriosis and diagnosed.
The record contained years of notes about weight and mental health. What it did not convey was the cost of years with an undiagnosed condition, the frustration of not being believed, or the loss of trust. Kate later described how different symptoms had repeatedly been attributed to her weight.
The point is not that every note must contain every detail of a patient’s life. Medical records have practical limits. But when an early explanation becomes part of the chart and the continuing symptoms or unanswered questions are less visible, later readers may inherit a distorted picture of the problem.
What AI can and cannot learn from a record
The study did not test an AI system. It did not measure whether AI would repeat the assumptions in these records or whether a particular AI tool would have reached a different diagnosis. The authors raise a concern about what could happen as AI systems rely on records that may already reflect bias, omissions, or delayed recognition.
That concern deserves attention. An AI tool that summarizes a chart may surface what is recorded most consistently. If the notes repeatedly emphasize weight or mental health while giving less attention to persistent pain and unanswered questions, a summary could preserve that imbalance. An AI scribe may capture more of a conversation, but producing a transcript is not the same as creating a complete, accurate clinical note. Choices about what to include and how to describe it still matter.
The same limitation applies to tools that search records for patterns. AI might help clinicians notice symptoms across visits or identify cases where a diagnosis was delayed. But that potential depends on the quality and context of the information available. A confident summary cannot turn missing information into evidence, and an AI-generated interpretation still needs a clinician to assess it.
A record is important, but it is not the whole story
Patients can sometimes help make important information easier to find. Before an appointment, it may be useful to write down:
- When the symptoms began and how they have changed
- How often they occur and how they affect daily activities
- What tests or treatments have already been tried, and what happened
- What remains unexplained and what questions still need an answer
During or after the visit, patients can ask whether the note reflects their main concern and what the next step is if the symptoms continue. They can also ask how to review their records and how to request a correction or add information if something important is missing. The process varies by health system.
These steps can support a clearer conversation, but they cannot put the responsibility for accurate records entirely on patients. Clinicians and health systems shape what is asked, interpreted, and documented. AI developers and health systems also need to examine whether their tools preserve uncertainty and patient-reported context, or simply make existing patterns easier to repeat.
The promise of AI in medicine should include more than reading records quickly. It should include helping clinicians notice patterns without mistaking the chart for the complete story, and helping patients’ concerns remain visible when earlier explanations have not resolved the problem.
Sources
