Where AI already works in healthcare — and where the risk is
The everyday uses aren't futuristic diagnostics; they're paperwork: visit summaries, discharge letters, insurer correspondence, translation and chart summarization. That is exactly where the risk sits — the input is a document carrying a name, ID, diagnoses and treatments, and the destination is an external server.
The numbers explain the sensitivity: per IBM Cost of a Data Breach 2024, a healthcare breach averages $9.77 million — the most expensive of all industries, for the 14th consecutive year. Medical data can't be "replaced" like a credit card.
What must never enter a public AI tool in a medical setting
- Identifying patient details — name, ID number, date of birth, address, phone.
- Diagnoses and treatments next to any identifier — "a 47-year-old patient from town X with diagnosis Y" identifies too.
- Scanned documents — referrals, summaries and photographed records carry identifiers on every line.
- Staff information — employee details, clinical-system passwords, permission structures.
- Non-de-identified research data — study files with participant identifiers.
How to still use AI on medical documents
- Set policy: which tools are approved, business-tier only, no training on your data.
- Require anonymization before every paste: the patient's name becomes PERSON_001, the ID becomes ID_001 — diagnoses and clinical course stay intact, so the AI's summary stays accurate.
- Prefer local processing: a tool that works in the browser or on the device — the original document never leaves the institution.
- Clean Word metadata — the Author field exposes the drafting clinician's name.
- Train the staff: most leaks are a visit summary pasted into a chat in good faith.
The principle matches the general list of what never to paste into ChatGPT — only the bar here is higher: with medical data, even a single leak is a reportable incident and real harm to a person.