
Stonewall did not create a population; it made an existing one impossible to ignore. Police records from the June 1969 uprising include charges of disorderly conduct and resisting arrest.1 Those records preserved an encounter with authority, not the fullness of the lives involved.
Consider Raymond Castro. One complaint report names him, age 27, among three people who “did shove and kick the officer.”1 It does not say that he had already been released from the raid when he saw a young friend still inside, without identification, signaling that he wanted out. “So naturally I tried to help him,” Castro later recalled.2
The record was not empty. It was filled in with the institution’s version of the night. Castro was not omitted. He was replaced by a partial truth that looked complete.
Our charts have done the same thing. LGBTQIA+ patients were never missing from the ICU. They were admitted, intubated, and discharged, and some died there. The encounters exist. What was erased is who they were and who they belonged to. A structured sex field may record sex assigned at birth rather than a patient’s gender.3 A partner can become “friend.” The next-of-kin line can hold a relative the patient would never have chosen.
Omission leaves a blank, and a blank announces itself. Substitution passes inspection. A database query sees a complete record and counts a different patient.
This is our algorithmic Stonewall moment. Artificial intelligence (AI) can now read the narrative, where the substitution often shows its seams. The same capability could expose people who never chose to be found.
That is the dilemma: Let AI read free text, under safeguards, to study care—not to decide who a patient is.
Hidden in plain sight
Imagine a patient who, before intubation, asks that her partner attend every family meeting. Now she is sedated and cannot repeat that request. Her partner sits at the bedside, knowing what she fears and what matters to her, but unsure whether anyone will ask. The structured record lists her mother as next of kin. Only a nursing note preserves the sentence that matters: “This is my person.”
This is where AI can help. Natural language processing means using computers to analyze written language; researchers have used it to detect stigmatizing language in clinical notes.3 For an ICU study, a model could extract the documented support person, the patient’s stated preference, and whether meeting participation was recorded. Reviewers would check each result against the original passage.
The distinction is crucial. A note about a partner does not establish sexual orientation. An algorithm should organize what was documented, preserve uncertainty, and never turn a clue into an identity. If it does, it writes a new substitution of its own.
What reading the notes can reveal
Researchers screened 331,794 notes from the Medical Information Mart for Intensive Care IV (MIMIC-IV) database and compared 366 clinician notes about patients who are gender-expansive with 388 comparison notes.3 Human reviewers found stigmatizing language in 62.3% vs 25.5% of notes for patients who were gender-expansive and nongender-expansive, respectively; the gap persisted when misgendering was excluded.3 These were selected, not representative, samples.
But the finding is the point: The narrative recorded these patients and too often described them in someone else’s terms.
Patients describe both sides of the risk. In one public discussion, participants welcomed chances to disclose gender identity but hesitated where it felt unsafe.4 Elsewhere, a patient described feeling alarmed when a former name resurfaced in a hospital portal.5 Being rewritten and being exposed are both real harms.
Start with an ICU question
Let’s return to the instance of a partner at the bedside. Was her request honored once she could no longer speak? The goal is not to classify the relationship but to learn whether the care provided respected the patient’s wishes.
For credentialed researchers, notes from the MIMIC-IV database provides discharge summaries linked to hospitalizations; access requires training and a data-use agreement.6 A starting query would select discharge-summary text with patient and admission identifiers (text, subject_id, hadm_id) and search for passages about family meetings, support people, and surrogates.7
AI could then extract requested and documented involvement, with supporting excerpts, for human review. Count admissions, not repeated mentions. Report “not documented” instead of “did not happen.”
Recognizing chosen family does not automatically confer legal authority. When a patient lacks capacity and has not appointed a surrogate, state laws vary in whom they authorize and prioritize.8 Clinicians should verify existing designations and seek ethics or legal guidance when authority is unclear.8 Honoring relationships and confirming legal authority go together.
What to do Monday
Start with a small, approved audit, not an unrestricted search for identities.
| Do | Do not |
| Define one care question and confirm permissions before extracting data | Treat access as permission for every possible use |
| Extract documented statements with context and supporting passages | Infer identity from names, medications, partners, or stereotypes |
| Validate results manually, including records the model missed | Treat model output or a filled-in field as ground truth |
| Report documentation gaps, uncertainty, and aggregate findings | Treat a selected cohort as population prevalence |
| Use an approved secure environment, restricted access, and small-cell protections | Upload restricted notes to public AI tools or publish identifiable excerpts |
| Involve affected communities in deciding purpose, access, and retention | Write inferred identities back into charts or override patient preferences |
Safeguards do not settle consent. Does documenting something for due diligence justify extracting it for research? Who decides, and what stays off limits? Those questions belong with patients and communities, not just researchers.
Raymond Castro’s report recorded the arrest and left out the friend. Our charts record the ventilator settings and leave out the person holding her partner’s hand. For the patient who cannot speak, a clinician’s promise is to provide care that remembers her truth.
Let the algorithm help us read what the record replaced. Never let it decide who the patient is.
This content was developed with assistance from artificial intelligence tools; it has been reviewed and edited by a human.
References
1. Fitzsimmons T. Stonewall police records shed new light on 1969 uprising. NBC News. June 28, 2019.
2. American Experience. Who was at Stonewall? PBS.
3. Xue L, Chayko M, Singh VK. Stigmatizing language in gender-expansive patient records: corpus development, disparity analysis, and natural language processing-based study. J Med Internet Res. 2026;28:e91089. doi:10.2196/91089
4. chondrichth_yes. Do you prefer to be asked gender identity at the doctor? Reddit, r/ftm.
5. udcvr. How did this hospital know my deadname? Reddit, r/FTMMen.
6. Johnson A, Pollard T, Horng S, Celi LA, Mark R. MIMIC-IV-Note: deidentified free-text clinical notes, Version 2.2. PhysioNet. Published January 6, 2023.
7. MIT Laboratory for Computational Physiology. Discharge. MIMIC-IV documentation.
8. Pope TM. Consent and surrogate decision making. Merck Manual Professional Edition.