Strip the PII before you paste into ChatGPT or Claude

AI has moved to the point of care

Clinicians now reach for AI in the middle of a visit, not just at a desk afterward. Purpose-built clinical assistants like OpenEvidence and UpToDate Expert AI sit alongside general assistants such as ChatGPT, Gemini and Claude, and the line between the two is blurring quickly.

A June 2026 study in Nature Medicine found that general-purpose models now match or beat specialized clinical tools on medical benchmarks. Gemini scored 97.4% on licensing-exam questions against 88 to 90% for the purpose-built systems, and on real clinician queries the frontier models were rated higher than the specialized tools. The practical takeaway for a busy clinician is blunt: the assistant that gives the best answer is no longer always the one labeled "medical", and the best answers come from giving the model the real case.

Every one of these tools is a third party

The catch is the same whichever you choose. OpenEvidence, UpToDate Expert AI, ChatGPT and Claude all run in the cloud. They produce sharper answers when you feed them the actual patient history, and the moment a name, Social Security number, medical record number or date of birth leaves your browser, you have disclosed identifiable patient data to an outside service.

"It is a clinical tool" is not the same as "it is safe to paste patient data into". Some vendors will sign a business associate agreement (BAA) for enterprise accounts; many clinicians use the free consumer tier, which carries no BAA. And because a general model now often gives the stronger answer, clinicians increasingly paste into tools that were never designed to hold health data at all.

De-identify in the browser, then ask

RedactProof removes the identifiers before anything is pasted. Load the letter, discharge summary or note, and on-device detection flags names, Social Security numbers, medical record numbers, dates of birth, addresses, contact details and 60+ other types of personal data. You redact them, copy the de-identified text, and paste only that into whichever assistant you trust. The file and the personal data in it never leave your device.

You keep what makes the tool useful, the clinical reasoning, the differential, the guideline lookup, without handing over anything that identifies the patient. The same workflow protects you whether you are using a clinical-native assistant or a general LLM, which matters now that the two are converging on quality.

What to strip before a clinical AI query

Direct identifiers. Patient name, Social Security number, medical record number, date of birth, address and ZIP code, phone and email. These are the obvious ones, and the ones most likely to be copied across without a second thought.

Quasi-identifiers. A rare diagnosis, an age and a small town can identify a patient as surely as a name. HIPAA Safe Harbor treats full dates and detailed geography as identifiers for a reason. Remove admission dates, unusual occupations and location detail wherever they are not clinically necessary to the question you are asking.

Third-party data. Names of relatives, caregivers, the referring physician and other clinicians mentioned in the record. They are personal data too, and they are not yours to disclose.

Free-text leakage. Notes and discharge summaries bury identifiers mid-sentence. Automated detection catches these far more consistently than a manual skim under time pressure, which is exactly when these queries tend to happen.

Not only for clinicians

The same problem turns up anywhere people paste source material into AI: a lawyer summarizing a brief, an HR team drafting from a case file, support staff pasting tickets. Redacting on-device first is the simple control that lets a team use AI without leaking the personal data inside its documents. If you work mainly with patient records, our healthcare redaction page covers de-identifying medical documents in more detail.