AI Finds Bias in Half of Pregnancy Notes

The number

A nurse writes "patient was combative and refused to cooperate." A resident writes "mother appeared calm and was a good historian." Same ward, same shift type, different patient. Multiply that pattern across a hospital system and you get a pattern in the data, not just a bad day for one clinician.

Researchers just measured how often that pattern shows up. A new study posted on medRxiv on July 28, 2026 ran a keyword-guided BERT classifier over 640,345 obstetric clinical notes covering 26,178 pregnancies at a single academic medical center. It flagged stigmatizing language in 47% of all pregnancies in the cohort. Not a fringe finding buried in a footnote. Nearly one in two.

Who it lands on

The disparity was not random. Black patients had significantly higher odds of stigmatizing documentation than Asian patients (adjusted odds ratio 1.5) or white patients (adjusted odds ratio 1.4). Patients with only a high school education were flagged more often than college-educated patients (aOR 1.5). Patients with preterm births, indicated (aOR 1.5) or spontaneous (aOR 1.2), were also more likely to carry stigmatizing language in their charts.

None of this required the classifier to know a patient's outcome. It just read what was written.

Why a language study matters to anyone building clinical AI

Here is the part that should worry anyone shipping decision support tools, not just anyone studying disparities. Every predictive model trained on hospital text learns from the note, not from the patient. If the note calls one mother "non-compliant" and another "cooperative" for the same behavior, a model trained on that corpus inherits the label, not the behavior. The bias does not stay contained in the sentence someone wrote at 2 a.m. It becomes a feature.

That is the quiet failure mode of "AI trained on your EHR data." The data was never neutral. It was written by tired people under time pressure, describing patients through whatever lens they carried into the room that day. A model does not correct for that. It compounds it, at scale, forever, unless someone goes looking.

The note is the record, and the record is now training data

We think about clinical notes as documentation: a record of what happened, written after the fact, mostly for billing and handover. This study is a reminder that the note has quietly become something else too. It is the substrate every downstream AI tool gets built on. Sepsis alerts, readmission risk scores, maternal health predictors, whatever comes next. All of it reads the chart, not the room.

That raises the bar for what "AI in the workflow" should mean. A tool that just autocompletes or summarizes whatever bias is already in the note makes the problem faster, not smaller. A tool worth trusting needs to be legible about where a claim in a note came from and traceable back to something a clinician actually observed, not phrasing inherited from the last shift's habits. That is a documentation problem before it is a model problem.

The researchers are not proposing a fix in this preprint, and neither are we. But the finding is a useful gut check for anyone who assumes more AI in the chart automatically means better care. Sometimes it just means the same bias, typed faster.

Sources

• Simha N, Takasuka H, Chen LC, Khan U, Oskotsky TT, Sirota M, Capra JA, Chen IY. "Forty-seven percent of pregnancies have stigmatizing language in their clinical notes in an electronic health record cohort." medRxiv, posted July 28, 2026. https://doi.org/10.64898/2026.07.27.26359049
• Yesil Science, "AI finds biased language in half of pregnancy records." https://yesilscience.com/ai-finds-biased-language-in-half-of-pregnancy-records/